WASTE SEGREGATION SYSTEM USING IMAGE

CLASSIFICATION AND DEEP LEARNING FOR SMART

CITIES-SmartWasteNet
Project submitted in partial fulfilment of the requirements for the award of the degree of

BACHELOR OF TECHNOLOGY
IN
CSE-DATA SCIENCE

KURELLA VISHNU PRIYA 22C91A6769

NEELI KEERTHI SATYA PRIYA 22C91A6797
YERROJU SOWMYA 22C91A67C6
UPPULETI BHAVANI 22C91A67C1
Under the Esteemed guidance of
Dr. A. SATISH KUMAR
Associate Professor
DEPARTMENT OF CSE-DATA SCIECNE

HOLY MARY INSTITUTE OF TECHNOLOGY AND SCIENCE

(Approved by AICTE New Delhi, Affiliated to JNTU Hyderabad, Accredited by NAAC with ‘A’ Grade)
Bogaram (V), Keesara (M), Medchal Dist-501301.
 HOLY MARY INSTITUTE OF TECHNOLOGY AND SCIENCE
(Approved by AICTE New Delhi, Affiliated to JNTU Hyderabad, Accredited by NAAC with ‘A’ Grade)

Bogaram (V), Keesara (M), Medchal Dist-501301.

DEPARTMENT OF CSE-DATA SCIENCE

CERTIFICATE
This is to certify that the major project entitled “WASTE SEGREGATION SYSTEM USING
IMAGE CLASSIFICATION AND DEEP LEARNING FOR SMART CITIES-
SmartWasteNet” is being submitted by KURELLA VISHNU PRIYA (22C91A6769), NEELI
KEERTHI SATYA PRIYA (22C91A6797), YERROJU SOWMYA (22C91A67C6),
UPPULETI BHAVANI (22C91A67C1) in partial fulfillment of the academic requirements for
the award of the degree of Bachelor of Technology in “CSE-DATA SCIENCE” from HOLY MARY
INSTITUTE OF TECHNOLOGY &SCIENCE,JNTU Hyderabad during the year 2022-2026.

Dr. A. Satish Kumar Dr. S. Srinivas

Associate Professor Associate Professor &HOD
[Link] CSE-DATA SCIENCE Dept of CSE-DATA SCIENCE
INTERNAL GUIDE HEAD OF THE DEPARTMENT

EXTERNAL EXAMINER
 ACKNOWLEDGEMENT

The satisfaction and euphoria that accompany the successful completion of any task would
be incomplete without the mention of the people who made it possible, whose constant
guidance and encouragement crown all efforts with success.

I take this opportunity to express my profound gratitude and deep regards to my guide Dr.A.
Sathish Kumar, Associate Professor, Dept. of CSE-DATA SCIENCE, Holy Mary
Institute of Technology, for his exemplary guidance, monitoring, and constant
encouragement throughout the project work.

My special thanks to Dr. S. Srinivas Head of the Department, CSE - Data Science, Holy
Mary Institute of Technology & Science, who has given immense support throughout the
course of the project.

I also thank Dr. P. Chandrasekhar, the Honourable Principal of my college, Holy Mary
Institute of Technology & Science, for providing me the opportunity to carry out this work.

At the outset, I express my deep sense of gratitude to the beloved Founder Chairman Dr.
A. Vara Prasad Reddy, Chairman A. Siddhartha Reddy, and Secretary Dr. A. Vijaya
Sarada Reddy of Holy Mary Institute of Technology & Science for giving me the
opportunity to complete my course of work.

I am obliged to the staff members of Holy Mary Institute of Technology & Science for the
valuable information provided by them in their respective fields. I am grateful for their
cooperation during the period of my assignment.

Last but not least, I thank Almighty, my Parents, and Friends for their constant
encouragement without which this assignment would not have been possible.

KURELLA VISHNU PRIYA 22C91A6769

NEELI KEERTHI SATYA PRIYA 22C91A6797
YERROJU SOWMYA 22C91A67C6
UPPULETI BHAVANI 22C91A67C1
 DECLARATION

This is to certify that the work reported in the present project titled “WASTE
SEGREGATION SYSTEM USING IMAGE CLASSIFICATION AND DEEP
LEARNING FOR SMART CITIES-SmartWasteNet” is a record of work done by us in
Department of CSE-DATA SCIECNE, Holy Mary Institute of Technology and Science.

To the best of my knowledge no part of the thesis is copied from books. Journals/internet and
wherever the portion is taken, the same has been duly referred to in the text. The reports are
based on the project work done entirely by us not copied from any other source.

KURELLA VISHNU PRIYA 22C91A6769

NEELI KEERTHI SATYA PRIYA 22C91A6797
YERROJU SOWMYA 22C91A67C6
UPPULETI BHAVANI 22C91A67C1
 ABSTRACT
Nowadays, automated waste segregation systems are crucial for sustainability and
urban hygiene, effective waste management has become a major concern for smart cities in the
modern era. In recent years, deep learning-based methods and image classification have
demonstrated potential for waste identification automation. However, traditional approaches,
like the base method described in Knowledge Based Systems, have drawbacks such as flat
classification pipelines, low robustness to obscured or cluttered waste images, and only
Convolution Neural Network (CNN) for feature extraction. Moreover, SmartWasteNet where
a hybrid deep learning framework that combines Transformer-based global context modelling
along with Deep Pyramid Convolutional Neural Network (DP-CNN) for multi-scale local
feature extraction and confidence-guided hierarchical classification strategy for adaptive
decision-making, to address these issues. Firstly, TACO (Trash Annotation in Context) and
TrashNet dataset which contain different waste images conducts conceptual preprocessing
using hierarchical labelling into coarse (organic vs. non-organic) and fine (plastic, metal, glass,
paper, hazardous, residual) categories. Furthermore, Discriminative representations are created
by fusing global and local features, and they are initially categorized at the coarse level.
Moreover, identification of waste types and low-confidence predictions are adaptively routed
to a fine-grained classifier. Furthermore, the Scalability and practical application are
guaranteed by software design. Additionally, the efficacy of SmartWasteNet for intelligent
urban waste segregation is demonstrated by experimental results that show a significant
improvement in accuracy, precision, recall, F1-score, robustness to occlusion, and adaptivity
when compared to the base method. Finally, adaptive hierarchical decision making is made by
this novel model. The primary drawback is the increased conceptual complexity brought due
to hierarchical architecture.

I
 TABLE OF CONTENTS

PAGE NO
CONTENTS

ABSTRACT I

1. INTRODUCTION 1-2
1.1 INTRODUCTION OF THE PROJECT 1
1.2 PROBLEM DEFINITION 2
1.3 OBJECTIVE OF THE PROJECT 2

2. LITERATURE SURVEY 3-4

3. SYSTEM ANALYSIS 5-10

3.1 EXISTING SYSTEM 5
3.2 PROPOSED SYSTEM 6
3.3 SYSTEM REQUIREMENTS 7
9
3.4 SYSTEM STUDY

4. SYSTEM DESIGN 11
4.1 SYSTEM ARCHITECTURE 11
4.2 DATAFLOW DIAGRAM 14
4.3 UML DIAGRAMS 15

5. SYSTEM IMPLEMENTATION 20
5.1 MODULES 20
5.2 ALGORITHM 22
5.3 TECHNOLOGIES 24
5.4 CODING 25

II
7. TESTING 35

8. OUTPUT SCREENS 39

9. CONCLUSION 41

10. FUTURE ENHANCEMENTS 42

11. REFERENCES 44

III
 1. INTRODUCTION

1.1 INTRODUCTION OF THE PROJECT:

Rapid urbanization and population growth have significantly increased municipal solid
waste, creating serious environmental, economic, and public health challenges in smart cities.
Efficient waste management is essential for sustainable urban development, as improper
disposal leads to soil pollution, greenhouse gas emissions, and resource depletion. Waste
segregation at the source improves recycling efficiency and reduces landfill waste, but manual
segregation is labor-intensive, time-consuming, and prone to errors. Recent advancements in
artificial intelligence, especially deep learning and image classification, have enabled
automated waste segregation systems. Convolutional Neural Networks (CNNs) can extract
visual features from waste images and classify them into categories such as organic, recyclable,
and hazardous. These systems enhance operational efficiency, reduce costs, and support
sustainable practices. However, challenges such as visual blockages, lighting variations,
background noise, and diverse waste shapes affect model performance. Deep learning models
also require large labeled datasets, which are difficult and costly to obtain. Additionally, high
computational and memory requirements limit deployment on low-power devices like smart
bins. Lack of generalization across regions due to variations in waste composition is another
issue. CNN models, though effective, may suffer from overfitting and limited accuracy on
complex data. Advanced models like ResNet improve accuracy using residual connections but
increase computational cost. Therefore, there is a need for scalable, accurate, and efficient
systems. Optimizing models for real-time deployment in IoT-based systems is crucial.
Automated waste classification can significantly improve recycling rates and reduce human
effort. Overall, intelligent waste segregation systems play a vital role in building cleaner and
more sustainable smart cities.

1
1.2 PROBLEM DEFINITION:

Rapid urbanization and population growth have significantly increased the volume of
municipal solid waste, creating serious environmental, economic, and public health challenges
in smart cities. The existing waste management systems rely heavily on manual segregation,
which is inefficient, time-consuming, and prone to human error. Traditional deep learning
approaches such as basic CNN models suffer from limitations like poor performance in complex
environments, lack of robustness to occlusion, lighting variations, and background clutter.
Furthermore, many existing systems require large labeled datasets and high computational
resources, making them difficult to deploy in real-world scenarios, especially in low-power
smart bins. The lack of adaptive decision-making and generalization across different regions
further reduces the effectiveness of current waste classification systems. These challenges
highlight the need for an efficient, scalable, and intelligent automated waste segregation system.

1.3 OBJECTIVE OF THE PROJECT

The objective of this project is to develop an intelligent waste segregation system,

SmartWasteNet, using advanced deep learning techniques for accurate and efficient
classification of waste. The system aims to automate waste classification by integrating Deep
Pyramid Convolutional Neural Networks (DP-CNN) for local feature extraction and
Transformer-based models for global context understanding. The system is designed to
implement a confidence-guided hierarchical classification strategy that enables adaptive
decision-making and improves classification accuracy. It also focuses on enhancing robustness
against real-world challenges such as occlusion, lighting variations, and complex backgrounds.
Additionally, the project aims to optimize model performance for real-time deployment in smart
city environments and IoT-enabled systems, thereby improving recycling efficiency, reducing
environmental impact, and supporting sustainable urban development.

2
 2. LITERATURE SURVEY

Megha Chhabra [Link] [6] established an intelligence waste classification approach based
on multi-layer using convolution neural network. Initially, data was collected from Kaggle,
which contained food and waste images. Subsequently, these images were preprocessed by
using normalization and data augmentation to reduce noise and ensure computational
complexity. Additionally, these images were fed into feature extraction using a convolution
layer which automatically extracts hierarchical spatial patterns between organic and recyclable
waste. Moreover, the automated CNN easily observes complex visual patterns without manual
features and leads to higher classification. However, CNN faced challenges in computational
cost due to scalability and real-world deployment.

Faizul Rakib Sayem et al. [7] introduced the DenseNet-201 framework for classifying
and detecting waste. Initially, the Waste Recycling Plant (WaRP) dataset consisted of 10,406
waste images. Subsequently, these images were preprocessed by using normalization, data
augmentation, and image resizing to improve model robustness. Additionally, these
preprocessed images were fed into dual-stream architecture combined with DenseNet-201,
enabling both local and global feature learning. Moreover, the DenseNet-201 method has high
classification accuracy and strong generalization in real recycling. However, the DenseNet-201
method failed in high computational complexity and increased training time.

Ahmet Alkilinc et al. [8] introduced the deep ensemble learning framework
combination of multiple pre-trained CNN models for effective waste classification. Initially ,
the TrashNet, TrashBox, Waste Pictures and Garbage Classification datasets is consisting of
labeled waste image data were used for training and evaluation. Then, these images were
preprocessed using image resizing, normalization, and data augmentation to improve the model
generalization. Subsequently, feature extraction was performed using pre-trained CNN models
such as DenseNet, ResNet, EfficientNet, and ConvNeXt where followed by averaging and
weighted ensemble strategies. Moreover, the pre-trained CNN models achieved the high
classification accuracy and also the strong robustness across multiple datasets. However, the

3
pre-trained CNN models suffer from high computational cost and increased inference
complexity.

Moshrof Hossian Dipo et al. [9] demonstrated a deep learning–based waste

classification system utilizing transfer learning for automatic waste recognition. Initially, a
public waste image dataset containing multiple waste categories, composed of RGB image data,
was employed. Subsequently, all these images were preprocessed using the image resizing,
normalization, and augmentation techniques to decrease overfitting. Furthermore, feature
extraction was having been carried out using pre-trained convolutional neural networks,
enabling effective learning of discriminative visual patterns. Moreover, the proposed approach
Yolo explained improved classification accuracy with reduced training effort. However, the
Yolo model showed limited performance in complex backgrounds and class-imbalanced
scenarios.

Md. Nahiduzzaman et al. [10] established a three-stage automated waste classification

system using a parallel lightweight depthwise separable Convolution Neural Network combined
with an ensemble Extreme Learning Machine (DP-CNN–En-ELM). Initially, the Tri Cascade
Waste Image dataset, consisting of 35,264 waste images, was constructed for multi-stage
classification. Then, these images were preprocessed by image resizing, pixel normalization
and scaling to optimize computational efficiency. Additionally, feature extraction has performed
by a custom lightweight DP-CNN architecture also followed by classification using the En-
ELM classifier. Moreover, the DP-CNN–En-ELM has gained high accuracy with low
computational overhead and real-time capability. However, the DP-CNN–En-ELM multi-stage
framework has introduced increased system complexity and implementation difficultl

4
 3. SYSTEM ANALYSIS

3.1 EXISTING SYSTEM

• Existing waste segregation systems in smart cities mainly rely on manual sorting and
traditional waste management techniques, which create multiple inefficiencies in large-scale
urban environments.

• Manual Segregation Dependency – Waste is separated by human workers, which is labor-

intensive, time-consuming, and prone to human errors, especially when handling large volumes
of waste.

• Limited Automation – Traditional systems use basic CNN models or flat classification
pipelines, which lack the ability to handle complex waste scenarios effectively.

• Poor Performance in Real Conditions – Existing models struggle with occlusion, background
clutter, lighting variations, and different waste shapes and textures, leading to reduced accuracy.

• High Computational Cost – Advanced models like ResNet and DenseNet provide better
accuracy but require high memory and computational resources, making them unsuitable for
real-time deployment in smart bins.

• Lack of Adaptability – Current systems follow non-adaptive classification, meaning they

cannot adjust decisions based on confidence levels or complexity of input data.

• Dataset Dependency – Most deep learning systems require large, well-labeled datasets, which
are costly and difficult to obtain in real-world scenarios.

• Limited Generalization – Existing models fail to perform consistently across different regions
due to variations in waste composition and packaging styles.

• Scalability Issues – Many systems are not scalable for smart city applications due to complex
architectures and lack of efficient deployment strategies.

5
3.2 PROPOSED SYSTEM

The proposed system, SmartWasteNet, is a hybrid deep learning-based intelligent waste

segregation framework designed for smart city environments. It improves classification
accuracy, adaptability, and efficiency by combining multiple advanced techniques.

Key Features and Functionality

• Hybrid Deep Learning Framework – Combines DP-CNN (Deep Pyramid CNN) for local
feature extraction and Transformer models for global context understanding.

• Multi-Scale Feature Extraction – Captures fine-grained textures, edges, and structural patterns
of waste using DP-CNN.

• Global Context Modeling – Transformer-based architecture captures long-range dependencies

and spatial relationships between objects in images.

• Feature Fusion Mechanism – Integrates local and global features into a unified representation
for improved classification accuracy.

• Confidence-Guided Hierarchical Classification – Performs classification in two stages:

• Coarse Level: Organic vs Non-Organic

• Fine Level: Plastic, Metal, Glass, Paper, Hazardous, Residual

• Adaptive Decision Making – Uses confidence scores to decide whether further classification
is required, reducing unnecessary computation.

• Robustness to Real-World Conditions – Handles occlusion, cluttered backgrounds, and

lighting variations effectively.

• Improved Performance Metrics – Achieves higher accuracy (96.1%), precision, recall, and
F1-score compared to traditional models.

6
• Scalable Architecture – Designed for integration with smart bins and IoT-based waste
management systems.

• Reduced Misclassification – Hierarchical approach minimizes errors in complex waste

scenarios.

3.3 SYSTEM REQUIREMENTS

SOFTWARE REQUIREMENTS

• Programming Language

• Python – Used for implementing deep learning models and data processing

• Deep Learning Frameworks

• TensorFlow / PyTorch – For building and training the SmartWasteNet model

• Image Processing

• OpenCV – For image preprocessing such as resizing and normalization

• Data Handling

• NumPy, Pandas – For efficient data manipulation and preprocessing

• Visualization Tools

• Matplotlib / Seaborn – For performance analysis and result visualization

• Development Environment

• Jupyter Notebook / VS Code – For coding, testing, and experimentation

7
• GPU Support

• CUDA-enabled libraries – For accelerating deep learning training

HARDWARE REQUIREMENTS

For Model Training

• Processor: High-performance CPU (Intel i7 / AMD Ryzen 7 or above)

• GPU: NVIDIA GPU (with CUDA support) for faster computation
• RAM: Minimum 16GB for handling large datasets
• Storage: 100GB+ SSD for datasets and model storage

For Deployment (Smart Systems / Edge Devices)

• Processor: Embedded processors or edge devices

• Memory: Optimized for low-power environments
• Device: Smart bins or IoT-enabled systems
• Network: Internet connectivity for data transmission

For End Users / Smart City Infrastructure

• Device: Monitoring systems, control dashboards

• Server: Cloud or centralized system for managing waste data
• Integration: IoT-enabled infrastructure for real-time waste tracking

8
3.4 SYSTEM STUDY

FEASIBILITY STUDY

The proposed SmartWasteNet system is evaluated to determine its feasibility in terms of

technical implementation, economic cost, operational efficiency, and practical deployment in
smart city environments. The system focuses on automated waste segregation using deep
learning techniques and intelligent decision-making.

• ECONOMICAL FEASIBILITY
• TECHNICAL FEASIBILITY
• OPERATIONAL FEASIBILITY
• PRACTICAL / IMPLEMENTATION ASPECTS

ECONOMICAL FEASIBILITY

The SmartWasteNet system is economically feasible as it mainly relies on software-based deep

learning models and publicly available datasets such as TACO and TrashNet. The primary costs
involve hardware setup (GPU systems), storage, and deployment infrastructure. Although
initial setup may require investment in computational resources, the system reduces long-term
labor costs by minimizing manual waste segregation. Additionally, improved waste
management leads to better recycling efficiency and reduced environmental costs, making it
cost-effective in the long run.

TECHNICAL FEASIBILITY

The system is technically feasible as it uses advanced yet widely available deep learning
technologies. It integrates DP-CNN for local feature extraction and Transformer models for
global context understanding, ensuring high classification performance. The implementation is
supported by standard tools such as Python, deep learning frameworks (TensorFlow/PyTorch),
and image processing techniques. The system can be deployed on GPU-enabled environments
for training and optimized for edge devices or smart bins for real-time applications. Despite

9
challenges like high computational requirements, optimization techniques make deployment
achievable.

OPERATIONAL FEASIBILITY

The SmartWasteNet system is operationally efficient as it automates waste segregation,

reducing human effort and errors. It can be integrated into smart city infrastructure and IoT-
enabled waste management systems for real-time operation. The system improves waste
classification accuracy, decision-making, and processing speed, leading to cleaner urban
environments. It also enhances scalability and adaptability, allowing it to function effectively
under different environmental conditions such as lighting variations and background clutter.
Minimal human intervention is required, making it highly practical for large-scale urban
deployment.

PRACTICAL / IMPLEMENTATION ASPECTS

The system is designed for real-world applicability, using datasets that represent both controlled
and real-world waste conditions. The hybrid architecture ensures robust performance even in
complex scenarios like occlusion and mixed waste. It supports scalable deployment in smart
bins and urban waste systems, though challenges such as model complexity and computational
cost must be managed. Future improvements can focus on model optimization for low-power
devices, enhancing real-time usability.

10
 4. SYSTEM DESIGN

4.1 SYSTEM ARCHITECTURE:

This system architecture diagram is custom-designed based on the methodology and

Figure 1 of your research paper. It provides a professional visual representation that ties your
concepts together for your documentation and presentation. The SmartWasteNet system
architecture is designed to provide a structured and efficient framework for automated waste
classification using advanced deep learning techniques. The architecture integrates multiple
components, including data input, preprocessing, feature extraction, feature fusion,
classification, and output generation, forming a complete end to-end pipeline. At the initial
stage, waste images are captured through smart bins or collected from benchmark datasets.
These images serve as the primary input to the system. The preprocessing module then prepares
the data by performing operations such as resizing, normalization, and augmentation to ensure
consistency and improve model performance. The core of the architecture lies in its hybrid
feature extraction mechanism. The system employs DP-CNN to capture fine-grained local
features such as edges, textures, and shapes, while the Transformer model captures global
contextual information and long-range dependencies. This dual approach ensures a
comprehensive understanding of complex waste images. The extracted features from both
models are then combined in the feature fusion module, which generates a unified
representation of the input data. This step is crucial for improving classification accuracy, as it
integrates complementary information from both local and global perspectives. Following
feature fusion, the classification module performs hierarchical classification. Initially, the
system conducts coarse classification to categorize waste into broad groups. Based on the
confidence level, the system may proceed to fine-grained classification for more detailed
categorization. This adaptive mechanism enhances efficiency and reduces unnecessary
computations. The final stage of the architecture is the output module, which provides the
classified waste category along with a confidence score. The results 19 can be displayed to users
or integrated with smart waste management systems for automated decision-making.
Additionally, the architecture supports scalability and flexibility, allowing deployment on both
cloud platforms and edge devices. The modular design ensures that individual components can
be updated or improved without affecting the entire system. Overall, the SmartWasteNet system
11
architecture provides a robust, scalable, and efficient framework for intelligent waste
segregation, making it highly suitable for real-world smart city applications. The architecture
also incorporates a feedback mechanism that allows the system to improve its performance over
time through continuous learning and model updates. Furthermore, the system is designed to
support real-time processing, enabling immediate classification and response in smart bin
applications. Error handling and validation mechanisms are included to ensure reliable system
operation and to prevent incorrect classifications due to poor input quality. The architecture
also facilitates data logging and monitoring, which helps in analyzing system performance and
identifying areas for improvement. Security considerations are addressed by ensuring safe data
transmission and controlled access to system components.

12
Architecture Key Concepts:

1. Preprocessing (Left): This covers the image resizing (224x224), normalization, and
data augmentation you specified to handle real-world image variability.
2. Hybrid Feature Extraction (Center): This is the core 'brain' of SmartWasteNet,
showing the parallel paths of local feature learning (using DP-CNN, simulated by the
ResNet backbone for edges and textures) and global context modeling (using the
Vision Transformer for spatial relationships).
3. Confidence-Guided Hierarchical Classification (Right): This section visualizes
your main innovation. The fused features go to the coarse classifier (Organic/Non-
Organic). A logic diamond then makes an Adaptive Routing decision based on your
0.85 confidence score and coarse label. If the result is confident Organic, it is
finalized; otherwise, it is routed to the Fine-grained Classifier for material sorting.

Fig 4.3 System Architecture

13
4.2 DATA FLOW DIAGRAM

The data flow within the SmartWasteNet system is designed to ensure smooth and efficient
processing of waste images from input to final classification. The system handles both real-
time and dataset-based inputs, ensuring flexibility in operation. Each stage of the data flow is
optimized to reduce processing time and improve accuracy. The use of parallel processing in
feature extraction enhances system performance by allowing simultaneous computation of
local and global features. The hierarchical classification process ensures that data is processed
efficiently by reducing unnecessary computations. Overall, the data flow design supports
scalability, accuracy, and real-time applicability. A Data Flow Diagram (DFD) is a graphical
representation that illustrates how data moves through a system and how it is processed at
different stages. In the 21 SmartWasteNet system, the DFD helps in understanding the flow of
waste image data from input to final classification output. It provides a clear visualization of
system processes, data storage, and interactions between different components. The DFD is
divided into multiple levels to represent the system in a structured manner. The Level 0
diagram gives a high-level overview of the system, while the Level 1 diagram provides detailed
insights into internal processes. This layered representation helps in simplifying complex
system workflows and improves understanding of system functionality. In addition to the basic
flow, the Level 0 diagram emphasizes the interaction between external entities and the
SmartWasteNet system. The system acts as a central processing unit that receives input data,
processes it, and produces classified outputs. The inclusion of datasets as an external entity
highlights the importance of training and validation in improving model performance. The
simplicity of the Level 0 diagram helps in understanding the overall functionality of the system
without focusing on internal complexities. The Level 1 diagram provides a deeper insight into
the internal processing stages of the system. Each process is designed to handle specific tasks,
ensuring efficient data transformation at every stage. The use of multiple processing layers
improves the system’s ability to handle complex inputs. Additionally, the integration of feature
fusion and hierarchical classification enhances decision-making accuracy. This detailed
representation helps in understanding how the system processes data step-by-step, making it
easier to analyze and optimize performance

14
Fig 4.2 Data Flow Diagrams

15
4.3 UML DIAGRAMS

4.3.1 USE CASE DIAGRAM

This diagram illustrates the interactions between users and the system, detailing the goals and
functionalities:

• Actors: User, System.

• Use Cases: Upload Waste Image, View Coarse Classification, Routing Decision (if
required), View Fine Classification (Routed), Finalize Organic.

Fig 4.3.1 Use-Case Diagram

16
4.3.2 CLASS DIAGRAM

• Hierarchical Model Components (Left): This groups the two feature extractors (ResNet
Backbone (DP-CNN) for local features and Vision Transformer for global context)
under a common FeatureExtractor abstract class. They feed into a common
FusionLayer.
• Adaptive Classifier (Center): This section models your core innovation. The
HierarchicalClassifier aggregates nested Coarse and FineGrained classifiers, complete
with the confidence_threshold attribute of 0.85, derived from Section 3.5.
• Backend & Preprocessing (Right): This details the PreprocessingModule and
FastAPIBackend, showing endpoints like postWasteImage and how they interact with
the model.

annotations are provided to explain derived details, such as the Combined Cross-Entropy loss
function and example output classifications. This single diagram provides a comprehensive
technical blueprint of your software design

Fig 4.3.2 Class Diagram

17
4.3.3 ACTIVITY DIAGRAM

The system contains a representation of its procedural flow through an activity diagram.
The SmartWasteNet workflow begins with image input, where waste images are provided to
the system. The process then splits into hybrid paths, enabling parallel processing of data. One
path utilizes a Transformer (Vision Transformer) to capture global spatial relationships and
long-range dependencies from the image. The system processes data through parallel paths to
enhance feature extraction and improve classification accuracy. Based on the extracted features,
the system performs classification, where initial decisions are made. Finally, the process leads
to fine-grained classification, categorizing waste into multiple classes such as organic and other
specific waste types. The activity diagram includes various stages such as input processing,
feature extraction, parallel computation, classification decisions, and final output states,
ensuring an efficient and structured workflow for intelligent waste segregation.

Fig 4.3.3 Activity Diagram

18
4.3.4 SEQUENCE DIAGRAM

Sequence diagrams show how different components of the SmartWasteNet system interact
with each other. The chronological order of operations is important as it represents how the
waste image is processed step by step. The sequence begins with the image input, where the
user or smart bin provides the waste image to the system. The system then forwards the image
to preprocessing, followed by feature extraction using DP-CNN and Transformer models
through parallel paths. After feature extraction, the system performs feature fusion and
classification, starting with coarse classification and proceeding to fine-grained classification if
required. Finally, the classified waste category is generated as output. The components in the
sequence diagram communicate through messages, ensuring smooth data flow and interaction
between different stages of the SmartWasteNet system.

Fig 4.3.4 Sequence Diagram

19
 5. SYSTEM IMPLEMENTATION

5.1 MODULES

The SmartWasteNet system is divided into multiple modules to ensure efficient waste
classification, improved accuracy, and real-time applicability in smart city environments. Each
module plays a crucial role in automating waste segregation using deep learning techniques.

1. Data Collection Module

• Collects waste images from datasets such as TACO and TrashNet.

• Ensures diversity in data including real-world and controlled waste images.
• Supports different waste categories like organic, recyclable, and hazardous.

2. Data Preprocessing Module

• Performs preprocessing techniques such as resizing, normalization, and data augmentation.

• Enhances image quality and reduces noise caused by lighting variations.
• Improves model robustness by simulating real-world conditions (rotation, flipping,
brightness).

3. Local Feature Extraction Module (DP-CNN)

• Extracts multi-scale local features such as edges, textures, and shapes.

• Uses Deep Pyramid CNN to capture fine-grained visual details.
• Handles variations in object size and structure effectively.

4. Global Feature Extraction Module (Transformer)

• Uses Transformer-based architecture to capture global context.

• Identifies spatial relationships and long-range dependencies in images.
• Improves understanding of complex and cluttered waste scenes.

20
5. Feature Fusion Module

• Combines local and global features into a unified feature vector.

• Enhances classification accuracy by integrating complementary information.
• Provides a strong representation for decision-making.

6. Hierarchical Classification Module

• Performs coarse classification (Organic / Non-Organic).

• Uses confidence score to decide further processing.
• If confidence is low, performs fine-grained classification into categories like plastic, metal,
glass, paper, hazardous, and residual.

7. Confidence Evaluation Module

• Calculates prediction confidence for classification decisions.

• Enables adaptive routing to improve accuracy.
• Reduces misclassification in complex scenarios.

8. Output Generation Module

• Displays final classified waste category.

• Provides confidence score for prediction reliability.
• Supports integration with smart bins and monitoring systems.

9. System Monitoring & Performance Module

• Evaluates system performance using metrics like Accuracy, Precision, Recall, and F1-score.
• Tracks model efficiency and robustness.
• Helps in improving system performance over time.

21
5.2 ALGORITHM

Algorithm for SmartWasteNet Waste Classification System:

Step 1: Data Input

1. Start
2. Receive waste image from dataset or smart bin

Step 2: Data Preprocessing

3. Resize image to standard size

4. Normalize pixel values
5. Apply data augmentation (rotation, flipping, brightness adjustment)

Step 3: Feature Extraction

6. Extract local features using DP-CNN

7. Extract global features using Transformer model

Step 4: Feature Fusion

8. Combine local and global features

9. Generate unified feature representation

Step 5: Coarse Classification

10. Classify image into Organic or Non-Organic

11. Calculate confidence score

Step 6: Adaptive Decision Making

12. If confidence ≥ threshold → finalize classification

13. Else → proceed to fine-grained classification

22
Step 7: Fine-Grained Classification

14. Classify into detailed categories:

• Plastic
• Metal
• Glass
• Paper
• Hazardous
• Residual

Step 8: Output Generation

15. Display final waste category

16. Provide confidence score

Step 9: Performance Evaluation

17. Evaluate using Accuracy, Precision, Recall, F1-score

Step 10: End Process

18. End

23
5.3 TECHNOLOGIES

1. Data Processing & Machine Learning

• Python – Core language for model development

• TensorFlow / PyTorch – Deep learning frameworks for training models

2. Image Processing

• OpenCV – Image preprocessing (resizing, normalization)

• NumPy – Efficient numerical computations

3. Deep Learning Architecture

• DP-CNN – For multi-scale local feature extraction

• Transformer (Vision Transformer) – For global context modeling

4. Data Handling & Analysis

• Pandas – Data manipulation and dataset handling

• Matplotlib / Seaborn – Visualization of results and performance metrics

5. Hardware & Acceleration

• NVIDIA GPU with CUDA – Accelerates model training

• Multi-core CPU – Supports preprocessing and data handling

6. Deployment & Integration

• Smart Bins / IoT Systems – For real-time waste classification

7. Evaluation Metrics

• Accuracy, Precision, Recall, F1-score – For performance analysis

24
5.4 CODING

[Link]

import tensorflow as tf

from [Link] import ImageDataGenerator

from keras import layers, models

import os

train_dir = "DATASET/TRAIN"

test_dir = "DATASET/TEST"

IMG_SIZE = (224, 224)

BATCH_SIZE = 32

train_datagen = ImageDataGenerator(rescale=1./255, rotation_range=20,

zoom_range=0.3, horizontal_flip=True)

test_datagen = ImageDataGenerator(rescale=1./255)

train_data = train_datagen.flow_from_directory(train_dir, target_size=IMG_SIZE,

batch_size=BATCH_SIZE, class_mode='binary')

test_data = test_datagen.flow_from_directory(test_dir, target_size=IMG_SIZE,

batch_size=BATCH_SIZE, class_mode='binary')

model = [Link]([

layers.Conv2D(32, (3,3), activation='relu', input_shape=(224,224,3)),

layers.MaxPooling2D(2,2),

25
 layers.Conv2D(64, (3,3), activation='relu'),

layers.MaxPooling2D(2,2),

layers.Conv2D(128, (3,3), activation='relu'),

layers.MaxPooling2D(2,2),

[Link](),

[Link](128, activation='relu'),

[Link](0.5),

[Link](1, activation='sigmoid')])

[Link](optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

[Link](train_data, validation_data=test_data, epochs=25)

[Link]("models", exist_ok=True)

[Link]("models/waste_model.h5")

print("Model saved!")

26
[Link]

from flask import Flask, render_template, request

from [Link] import load_model

from [Link] import image

import numpy as np

import os

app = Flask(__name__)

model = load_model("models/waste_model.h5")

UPLOAD_FOLDER = "static/uploads"

# Create folder automatically if not exists

[Link](UPLOAD_FOLDER, exist_ok=True)

[Link]["UPLOAD_FOLDER"] = UPLOAD_FOLDER

def predict_label(img_path):

img = image.load_img(img_path, target_size=(224, 224))

img_array = image.img_to_array(img) / 255.0

img_array = np.expand_dims(img_array, axis=0)

pred = [Link](img_array)[0][0]

return "Inorganic Waste " if pred > 0.7 else "Organic Waste "

@[Link]("/")

27
def index():

return render_template("[Link]")

@[Link]("/predict", methods=["POST"])

def predict():

file = [Link]["file"]

if file:

filepath = [Link]([Link]["UPLOAD_FOLDER"], [Link])

[Link](filepath)

result = predict_label(filepath)

return
render_template("[Link]",prediction=result,image_path="static/uploads/" +
[Link])

return "No file uploaded"

if __name__ == "__main__":

[Link](debug=True)

28
[Link]

<!DOCTYPE html>

<html lang="en">

<head>

<meta charset="UTF-8">

<meta name="viewport" content="width=device-width, initial-scale=1.0">

<title>SmartWasteNet</title>

<link rel="stylesheet" href="{{ url_for('static', filename='css/[Link]') }}">

</head>

<body>

<div class="container">

<h1>♻ Smart Waste Classification</h1>

<p>Upload an image to classify waste as Organic or Inorganic</p>

<form action="/predict" method="post" enctype="multipart/form-data"

class="upload-box">

<input type="file" name="file" required>

<button type="submit">Predict</button>

</form>

</div>

29
 </body>

</html>

[Link]

<!DOCTYPE html>

<html lang="en">

<head>

<meta charset="UTF-8">

<meta name="viewport" content="width=device-width, initial-scale=1.0">

<title>Result</title>

<link rel="stylesheet" href="{{ url_for('static', filename='css/[Link]') }}">

</head>

<body>

<div class="container">

<h1>Prediction Result</h1>

<img src="{{ image_path }}" class="result-img">

<h2 class="result-text">{{ prediction }}</h2>

<a href="/" class="btn">Try Another</a>

</div>

30
 </body>

</html>

[Link]

*{

margin: 0;

padding: 0;

box-sizing: border-box;

font-family: Arial, Helvetica, sans-serif;

body {

background: linear-gradient(135deg, #4CAF50, #2E7D32);

color: white;

display: flex;

justify-content: center;

align-items: center;

height: 100vh;

31
/* Container */

.container {

background: white;

color: black;

padding: 30px;

border-radius: 12px;

width: 90%;

max-width: 400px;

text-align: center;

box-shadow: 0 10px 25px rgba(0,0,0,0.2);

/* Title */

h1 {

margin-bottom: 10px;

/* Upload box */

.upload-box {

32
 margin-top: 20px;

.upload-box input {

width: 100%;

padding: 10px;

margin-bottom: 15px;

/* Button */

button, .btn {

background: #4CAF50;

color: white;

padding: 12px;

border: none;

width: 100%;

border-radius: 6px;

cursor: pointer;

font-size: 16px;

text-decoration: none;

display: inline-block;

33
}

utton:hover, .btn:hover {

background: #388E3C;

/* Result image */

.result-img {

width: 100%;

margin: 20px 0;

border-radius: 10px;

/* Result text */

.result-text { font-size: 22px;font-weight: bold;color: #2E7D32;}

/* Mobile responsive */

@media (max-width: 600px) {

.container {

padding: 20px;

34
 7. TESTING

1. Introduction to Testing

Testing is an essential phase in the development of the SmartWasteNet system to ensure

accurate waste classification and reliable performance in real-world environments. Various
testing methodologies are applied to validate the system’s functionality, robustness, efficiency,
and scalability. The testing process ensures that the deep learning model performs effectively
under different conditions such as lighting variations, occlusion, and background clutter.

2. Testing Strategies

To ensure the reliability and effectiveness of the SmartWasteNet system, the following testing
strategies were implemented:

a) Unit Testing

Unit testing was performed to validate individual components such as data preprocessing,
feature extraction (DP-CNN and Transformer), feature fusion, and classification modules. Each
module was tested independently to ensure correct functionality.

b) Integration Testing

Integration testing was conducted to verify the interaction between modules such as
preprocessing → feature extraction → classification. It ensures smooth data flow between
different stages of the system

c) System Testing

System testing evaluated the complete SmartWasteNet workflow, including image input,
preprocessing, feature extraction, hierarchical classification, and output generation. This
ensured that the entire system works correctly as an integrated unit.

35
d) Model Validation Testing

Model validation was performed using datasets such as TACO and TrashNet to evaluate
accuracy and generalization. The system was tested under different conditions like cluttered
backgrounds and varying lighting.

e) Performance Testing

Performance testing was carried out to evaluate system efficiency in terms of processing speed,
accuracy, and computational cost.

36
3. Test Cases

The system's multiple features went through testing using developed test cases. An
illustration test case table follows below.

Test Case Description Input Expected Actual Status

ID Output Output
TC-01 Image Image
Image Input Waste Image Pass
Accepted Accepted
TC-02 Processed Processed
Preprocessing Raw Image Pass
Image Image
TC-03 Feature Extracted Extracted
Image Data Pass
Extraction Features Features
TC-04 Coarse Organic/Non- Correct
Image Pass
Classification Organic Class
TC-05 Low
Fine Detailed Correct
Confidence Pass
Classification Category Category
Input
TC-06 Performance Model Accuracy > Achieved
Pass
Metrics Output 90% 96.1%

37
4. Bugs and Fixes

• Issue: Misclassification in complex backgrounds

o Fix: Improved feature extraction using Transformer-based global context modeling

• Issue: Low confidence in certain waste categories

o Fix: Introduced confidence-guided hierarchical classification

• Issue: High computational complexity

o Fix: Optimized model architecture and feature fusion techniques

5. Performance Testing

Performance testing was conducted to evaluate system efficiency:

• Accuracy Testing: Achieved high accuracy of 96.1%, outperforming traditional CNN

models

• Precision & Recall Testing: Ensured balanced classification performance

• Robustness Testing: Verified performance under occlusion, lighting variation, and clutter

• Computational Efficiency: Evaluated processing speed for real-time applicability

6. Security & Reliability Testing

Although primarily a deep learning system, reliability testing was performed:

• Data Integrity Testing: Ensured correct preprocessing and feature extraction

• Model Stability Testing: Verified consistent outputs across multiple inputs

• Error Handling: Ensured system handles low-confidence predictions using adaptive routing

• Scalability Testing: Evaluated performance for large datasets and real-world deploymen

38
 8. OUTPUT SCREENS

Figure 1: Front page

Figure 2 : Predicted Result as Inorganic Waste

39
Figure 3 : Predicted Result as Organic Waste

40
 [Link]

This research presented SmartWasteNet which is an intelligent deep learning–based

waste segregation framework fitted for smart city environments. Moreover, the proposed model
integrates Deep Pyramid Convolutional Neural Networks to capture multiscale local visual
features and Transformer-based global context modeling to learn long-range dependencies
within complex waste scenarios. Further, a confidence-guided hierarchical classification
strategy was performed to allow adaptive coarse-to-fine waste categorization, efficiently
addressing the disadvantages of flat CNN-based pipelines, such as reduced robustness under
occlusion and visual ambiguity. Experimental evaluations on benchmark waste image datasets
demonstrated that the proposed framework achieves consistent improvements in classification
reliability, accuracy, and adaptability when compared with existing approaches. The system
also shows enhanced performance in handling real-world challenges such as background
clutter, lighting variations, and diverse waste textures. In fact, although the hierarchical
architecture introduces additional conceptual complexity, it significantly improves decision-
making efficiency, scalability, and robustness for real-world urban deployments. Additionally,
the hybrid combination of local and global feature extraction enables the model to achieve a
better understanding of complex waste scenes, thereby reducing misclassification and
improving overall system performance. The proposed framework also supports integration with
smart city infrastructure, enabling automated waste management and contributing to sustainable
urban development. Moreover, future work will focus on optimizing model efficiency, reducing
computational cost, and simplifying the architecture to support real-time deployment on low-
power embedded systems such as smart bins. Further enhancements may include expanding
datasets, improving generalization across regions, and integrating IoT-based monitoring
systems for large-scale implementation. Overall, SmartWasteNet provides a scalable, accurate,
and intelligent solution for automated waste segregation, playing a significant role in improving
recycling efficiency, reducing environmental impact, and supporting the development of
cleaner and more sustainable smart cities.

41
 10. FUTURE ENHANCEMENTS

The SmartWasteNet system can be further improved by incorporating advanced

technologies and optimization techniques to enhance accuracy, efficiency, and real-time
deployment in smart city environments. Some potential future developments are:

1. Model Optimization for Edge Devices

• Reduce computational complexity to enable deployment on low-power smart bins and

embedded systems.
• Improve real-time performance with lightweight architectures.

2. Real-Time Waste Classification

• Implement real-time image processing for instant waste detection and segregation.
• Integrate with IoT-enabled smart bins for automated waste handling.

3. Improved Dataset Expansion

• Use larger and more diverse datasets to improve model generalization across different
regions.
• Include more real-world waste scenarios with complex backgrounds.

4. Advanced Hybrid Models

• Enhance the existing architecture by combining more efficient CNN and Transformer
variants.
• Improve accuracy and robustness in challenging conditions like occlusion and clutter.

5. Lightweight Deep Learning Models

• Develop compact models to reduce memory usage and training time.

• Make the system more suitable for large-scale deployment.

6. Multi-Class and Fine-Grained Classification Enhancement

42
• Improve classification accuracy for similar-looking waste categories.
• Add more detailed waste categories for better recycling efficiency.

7. Integration with Smart City Infrastructure

• Connect the system with IoT networks, smart bins, and centralized monitoring systems.
• Enable automated waste collection and management.

8. Automated Decision-Making Systems

• Improve the confidence-guided hierarchical classification mechanism.

• Enable smarter and more adaptive decision-making.

9. Energy-Efficient Processing

• Optimize models to reduce energy consumption in embedded devices.

• Support sustainable and eco-friendly deployment.

10. Cloud-Based Monitoring & Analytics

• Use cloud platforms for data storage, monitoring, and performance analysis.
• Provide real-time insights into waste management systems.

11. Explainable AI (XAI) Integration

• Add explain ability features to understand how the model makes decisions.
• Increase transparency and trust in automated systems.

43
 [Link]

[1] Celik, G., 2025. Multi-layer feature fusion for high-accuracy solid waste classification using
a hybrid deep learning model. The Visual Computer, pp.1-23.

[2] Castro-Bello, M., Roman-Padilla, D.B., Morales-Morales, C., Campos-Francisco, W.,

Marmolejo-Vega, C.V., Marmolejo-Duarte, C., Evangelista-Alcocer, Y.

and Gutiérrez-Valencia, D.E., 2025. Convolutional neural network models in municipal

solid waste classification: towards sustainable management. Sustainability, 17(8), p.3523.

[3] Gude, D.K., Bandari, H., Challa, A.K.R., Tasneem, S., Tasneem, Z., Bhattacharjee, S.B.,
Lalit, M., Flores, M.A.L. and Goyal, N., 2024. Transforming urban sanitation: enhancing
sustainability through machine learning-driven waste processing. Sustainability, 16(17),
p.7626.

[4] Chauhan, R., Shighra, S., Madkhali, H., Nguyen, L. and Prasad, M., 2023. Efficient future
waste management: A learning-based approach with deep neural networks for smart system
(LADS). Applied Sciences, 13(7), p.4140.

[5] Malik, M., Sharma, S., Uddin, M., Chen, C.L., Wu, C.M., Soni, P. and Chaudhary, S., 2022.
Waste classification for sustainable development using image recognition with deep
learning neural network models. Sustainability, 14(12), p.7222.

[6] Chhabra, M., Sharan, B., Elbarachi, M. and Kumar, M., 2024. Intelligent waste
classification approach based on improved multi-layered convolutional neural
network. Multimedia Tools and Applications, 83(36), pp.84095-84120.

[7] Sayem, F.R., Islam, M.S.B., Naznine, M., Nashbat, M., Hasan-Zia, M., Kunju, A.K.A.,
Khandakar, A., Ashraf, A., Majid, M.E., Kashem, S.B.A. and Chowdhury, M.E., 2025.
Enhancing waste sorting and recycling efficiency: robust deep learning-based approach for
classification and detection. Neural Computing and Applications, 37(6), pp.4567-4583.

44
[8] Alkılınç, A., Okay, F.Y., Kök, İ. and Özdemir, S., 2025. Deep Ensemble Learning Model
for Waste Classification Systems. Sustainability, 18(1), p.24.

[9] Dipo, M.H., Farid, F.A., Mahmud, M.S.A., Momtaz, M., Rahman, S., Uddin, J. and Karim,
H.A., 2025. Real-Time Waste Detection and Classification Using YOLOv12-Based Deep
Learning Model. Digital, 5(2), p.19.

[10] Nahiduzzaman, M., Ahamed, M.F., Naznine, M., Karim, M.J., Kibria, H.B., Ayari,
M.A., Khandakar, A., Ashraf, A., Ahsan, M. and Haider, J., 2025. An automated waste
classification system using deep learning techniques: Toward efficient waste recycling and
environmental sustainability. Knowledge-Based Systems, 310, p.113028.

[11] Gibellini, F., Fraternali, P., Boracchi, G., Morandini, L., Martinoli, T., Diecidue, A. and
Malegori, S., 2025. A deep learning pipeline for solid waste detection in remote sensing
images. Waste Management Bulletin, p.100246.

[12] Cheema, S.M., Hannan, A. and Pires, I.M., 2022. Smart waste management and
classification systems using cutting edge approach. Sustainability, 14(16), p.10226.

[13] Wang, Z., Zhou, W. and Li, Y., 2024. GFN: a garbage classification fusion network
incorporating multiple attention mechanisms. Electronics, 14(1), p.75.

[14] Nnamoko, N., Barrowclough, J. and Procter, J., 2022. Solid waste image classification
using deep convolutional neural network. Infrastructures, 7(4), p.47.

[15] Naznine, M., Nahiduzzaman, M., Karim, M.J., Ahamed, M.F., Salam, A., Ayari, M.A.,
Khandakar, A., Ashraf, A., Ahsan, M. and Haider, J., 2025. PLDs-CNNridge-ELM:
Interpretable lightweight waste classification framework. Engineering Applications of
Artificial Intelligence, 162, p.112522.

45
 14 II February 2026

[Link]
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 14 Issue II Feb 2026- Available at [Link]

Waste Segregation System using Image

Classification and Deep Learning for Smart Cities -
SmartWasteNet
Kurella Vishnu Priya1, Neeli Keerthi Satya Priya2, Yerroju Sowmya3, Uppuleti Bhavani4, Dr. A. Sathish Kumar5,
Dr. B. Venkataramana6
¹Student, BTech CSE(DS) 4th Year, Holy Mary Inst. of Tech. and Science, Hyderabad, TG, India,
vishnupriyakurella@[Link]
²Student, BTech CSE(DS) 4th Year, Holy Mary Inst. of Tech. and Science, Hyderabad, TG, India,
nkeerthisatyapriya@[Link]
³Student, BTech CSE(DS) 4th Year, Holy Mary Inst. of Tech. and Science, Hyderabad, TG, India,
yerrojusowmya27@[Link]
⁴Student, BTech CSE(DS) 4th Year, Holy Mary Inst. of Tech. and Science, Hyderabad, TG, India,
bhavaniuppuleti08@[Link]
⁵Assoc. prof, CSE(DS),Holy Mary Inst. of Tech. and Science, Hyderabad, TG, India,
sathishk0711@[Link]
⁶Assoc. prof, CSE(DS), Holy Mary Inst. of Tech. and Science, Hyderabad, TG, India,
venkataramana.b@[Link]

Abstract: Nowadays, automated waste segregation systems are crucial for sustainability and urban hygiene, effective waste
management has become a major concern for smart cities in the modern era. In recent years, deep learning-based methods and
image classification have demonstrated potential for waste identification automation. However, traditional approaches, like the
base method described in Knowledge Based Systems, have drawbacks such as flat classification pipelines, low robustness to
obscured or cluttered waste images, and only Convolution Neural Network (CNN) for feature extraction. Moreover,
SmartWasteNet where a hybrid deep learning framework that combines Transformer-based global context modelling along with
Deep Pyramid Convolutional Neural Network (DP-CNN) for multi-scale local feature extraction and confidence-guided
hierarchical classification strategy for adaptive decision-making, to address these issues. Firstly, TACO (Trash Annotation in
Context) and TrashNet dataset which contain different waste images conducts conceptual preprocessing using hierarchical
labelling into coarse (organic vs. non-organic) and fine (plastic, metal, glass, paper, hazardous, residual) categories.
Furthermore, Discriminative representations are created by fusing global and local features, and they are initially categorized at
the coarse level. Moreover, identification of waste types and low-confidence predictions are adaptively routed to a fine-grained
classifier. Furthermore, the Scalability and practical application are guaranteed by software design. Additionally, the efficacy of
SmartWasteNet for intelligent urban waste segregation is demonstrated by experimental results that show a significant
improvement in accuracy, precision, recall, F1-score, robustness to occlusion, and adaptivity when compared to the base method.
Finally, adaptive hierarchical decision making is made by this novel model. The primary drawback is the increased conceptual
complexity brought due to hierarchical architecture.
Keywords: SmartWasteNet, DP-CNN, Transformer, Confidence-guided hierarchical, fine-grained, coars

I. INTRODUCTION
In recent times, Rapid urbanization and population growth have significantly Boosted the volume of municipal solid waste, creating
Harsh environmental, economic, and public health challenges for smart cities. Moreover, Efficient waste management is a key
component of sustainable urban development, as improper disposal will impact to Soil pollution, greenhouse gas emissions, and
resource depletion. Waste segregation at the source plays a vital role in improving recycling efficiency and reducing Excess landfill
waste. However, manual segregation requires intensive labor, more time consuming, and leads to human error, especially in large-
scale urban environments [1]. Moreover, Recent development in artificial intelligence, particularly image classification and deep

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 240
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 14 Issue II Feb 2026- Available at [Link]

learning, has improvised to enable automated waste segregation systems capable of identifying, and classifying waste into
categories like organic, recyclable, and hazardous. Moreover, Deep learning models which are Convolutional Neural Networks
(CNNs) that automatically extract appropriate visual features from waste images and obtain high classification accuracy.
Furthermore, integrating such intelligent waste segregation systems into smart city infrastructure enhance operational efficiency led
to reduce costs, and support environmentally sustainable practices. However, the progress in automated waste segregation systems
face multiple challenges in Visual Blockages, variations lighting conditions, background clutter, different waste shapes and textures
often lead to degrade classification performance. Moreover, a lot of deep learning models require large, well labeled datasets for
training which are costly and complex to obtain in real-world waste management scenarios. [2] Additionally, high computational
complexity and memory requirements of deep neural networks restrict deployment on low power embedded systems which are
commonly used in smart bins.
Moreover, the lack of generalization across regions as waste composition and packaging methods or style may significantly vary
among countries. In Additional to the work there arises the immediate need to develop a scalable, accurate, and efficient waste
segregation system suitable for smart city applications. Furthermore, automating waste classification uses image-based deep
learning techniques to improve recycling rates and reduce human actions. Moreover, optimizing models balance accuracy and
computational efficiency to enable real-time deployment in smart bins and IoT-enabled waste collection systems [3]. In addition to
existing challenges where systems play an important role in cleaner cities where there is reduced environmental impact and more
sustainable urban living. Further, the State-of-the-Art methods followed by a basic Convolution Neural Network (CNN) model are
mostly used for waste image classification due to its ability to automatically learn structural features from images without manual
feature extraction. Moreover, CNNs have strong performance in visual recognition tasks and their relatively simple architecture
compared to deeper models. However, CNNs often suffer from limited accuracy when dealing with complex waste images and may
overfit when trained on small datasets [4]. Subsequently, ResNet is employed in waste classification tasks to overcome the
vanishing gradient problem by using residual connections which enable very deep network architectures. Moreover, improved
classification accuracy and better feature representation for complex and various stages of waste categories. However, ResNet has
high computational cost and memory usage, which limits its suitability for real time and edge device deployment [5].

II. LITERATURE
Megha Chhabra [Link] [6] established an intelligence waste classification approach based on multi-layer using convolution neural
network. Initially, data was collected from Kaggle, which contained food and waste images. Subsequently, these images were
preprocessed by using normalization and data augmentation to reduce noise and ensure computational complexity. Additionally,
these images were fed into feature extraction using a convolution layer which automatically extracts hierarchical spatial patterns
between organic and recyclable waste. Moreover, the automated CNN easily observes complex visual patterns without manual
features and leads to higher classification. However, CNN faced challenges in computational cost due to scalability and real-world
deployment.
Faizul Rakib Sayem et al. [7] introduced the DenseNet-201 framework for classifying and detecting waste. Initially, the Waste
Recycling Plant (WaRP) dataset consisted of 10,406 waste images. Subsequently, these images were preprocessed by using
normalization, data augmentation, and image resizing to improve model robustness. Additionally, these preprocessed images were
fed into dual-stream architecture combined with DenseNet-201, enabling both local and global feature learning. Moreover, the
DenseNet-201 method has high classification accuracy and strong generalization in real recycling. However, the DenseNet-201
method failed in high computational complexity and increased training time.
Ahmet Alkilinc et al. [8] introduced the deep ensemble learning framework combination of multiple pre-trained CNN models for
effective waste classification. Initially , the TrashNet, TrashBox, Waste Pictures and Garbage Classification datasets is consisting of
labeled waste image data were used for training and evaluation. Then, these images were preprocessed using image resizing,
normalization, and data augmentation to improve the model generalization. Subsequently, feature extraction was performed using
pre-trained CNN models such as DenseNet, ResNet, EfficientNet, and ConvNeXt where followed by averaging and weighted
ensemble strategies. Moreover, the pre-trained CNN models achieved the high classification accuracy and also the strong robustness
across multiple datasets. However, the pre-trained CNN models suffer from high computational cost and increased inference
complexity.
Moshrof Hossian Dipo et al. [9] demonstrated a deep learning–based waste classification system utilizing transfer learning for
automatic waste recognition. Initially, a public waste image dataset containing multiple waste categories, composed of RGB image
data, was employed. Subsequently, all these images were preprocessed using the image resizing, normalization, and augmentation

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 241
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 14 Issue II Feb 2026- Available at [Link]

techniques to decrease overfitting. Furthermore, feature extraction was has been carried out using pre-trained convolutional neural
networks, enabling effective learning of discriminative visual patterns. Moreover, the proposed approach Yolo explained improved
classification accuracy with reduced training effort. However, the YolO model showed limited performance in complex backgrounds
and class-imbalanced scenarios.
Md. Nahiduzzaman et al. [10] established a three-stage automated waste classification system using a parallel lightweight depth-
wise separable Convolution Neural Network combined with an ensemble Extreme Learning Machine (DP-CNN–En-ELM). Initially,
the Tri Cascade WasteImage dataset, consisting of 35,264 waste images, was constructed for multi-stage classification. Then, these
images were preprocessed by image resizing, pixel normalization and scaling to optimize computational efficiency. Additionally,
feature extraction has performed by a custom lightweight DP-CNN architecture also followed by classification using the En-ELM
classifier. Moreover, the DP-CNN–En-ELM has gained high accuracy with low computational overhead and real-time capability.
However, the DP-CNN–En-ELM multi-stage framework has introduced increased system complexity and implementation
difficultly.
Gaffari Celik et al. [11] introduced multi-layer feature fusion for solid waste classification and detection of waste segregation.
Initially, the Household-Garbage and TrashNet dataset were considered waste images. Then, these images were pre-processed by
using image resizing, normalization, and data augmentation to enhance the model performance. Subsequently, these pre-processed
images were fed into EfficientNetB0, InceptionV3 and HypeColumn for enabling the extraction of both general and detailed
features. Moreover, the XGBClassifier was performed in the classification and detection by optimizing performance. However, this
XGBClassifier suffered from high computational cost and increased inference complexity.
Sehrish Munawar Cheema et al. [12] presented a deep learning–based waste classification framework by utilizing the Convolutional
Neural Network (CNN) to enhance solid waste management system. Initially, a large-scale labeled Waste Image dataset considering
multiple waste categories.
Subsequently, the images were preprocessed by using image resizing, normalization, and data augmentation techniques to enrich the
model significantly and reduce overfitting. These preprocessed images were then fed into CNN-based architecture, which effectively
extracts texture-based and spatial features. Moreover, Smart Waste Management and Classification Mechanism using a Cutting-edge
Approach (SWMACM-CA) illustrated high classification accuracy and generalized performance under real-time situations.
However, the model exhibited increased training complexity and high computational cost.
Zhaoqi Wang et al. [13] implemented Deep learning–based waste classification framework using a convolutional neural network
(CNN) improve automated solid waste sorting. Initially, A large scale waste image dataset is collected and differentiate into multiple
recyclable and non-recyclable classes. Subsequently, the images were pre-processed using normalization, data augmentation, and
resizing techniques to improve feature representation and reduce overfitting. The pre-processed images are trained using a CNN
architecture with transfer learning to extract discriminative visual features. Moreover, the CNN enhanced high classification
accuracy and improved robustness under different lighting and background conditions. However, the model suffered from increased
computational cost and required high memory resources limiting its deployment on low-power embedded systems.
Nonso Nnamoko et al. [14] illustrated an Intelligent Waste Segregation System using Deep Learning techniques for smart
infrastructure applications.
Initially, a labeled waste image dataset was collected from real-world conditions. Subsequently, these images are preprocessed by
using image scaling, noise removal and normalization techniques to improve data quality. Additionally, the processed images were
input into a Deep Convolutional Neural Network (DCNN), which permits the model to observe reliable features for waste
classification. Furthermore, the DCNN method established high robustness and reliable concert in waste segregation tasks.
However, the system demonstrated challenges in scalability, and this is essential for significant computational resources for training
and deployment of waste segregation.
Mansura Naznine et al. [15] introduced the lightweight deep learning framework for real-time waste classification which aims at the
smart city applications. Initially, waste images were collected from publicly available datasets and resized to standard dimensions.
Then, applied into pre-process using Data augmentation and normalization techniques which has applied to improve robustness
against environmental variations. Additionally, the processed images were then trained using a compact CNN architecture optimized
for efficiency. Moreover, the Parallel Depth-wise Separable Convolutional Neural Network (PLDs-CNN) model has achieved
acceptable accuracy with faster inference speed making it suitable for real-world deployment. However, the lightweight design were
resulted in reduced performance when classifying complex or overlapping waste objects.

III. METHODOLOGY

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 242
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 14 Issue II Feb 2026- Available at [Link]

This research proposed a hybrid deep learning-based waste segregation method called SmartWasteNet for smart city applications.
Firstly, data is collected which contains waste images from datasets like TACO and TrashNet. Moreover, due to differences in
lighting, scale, and background, data is preprocessed by resizing, normalizing, and using data augmentation techniques. Further,
they used a DP-CNN method to extract multi-scale local features while a Transformer-based method captured global contextual
information. Finally, they applied a confidence-guided hierarchical classification method to differentiate the waste into organic and
non-organic classes. Furthermore, refining the samples into more specific categories. The overall methodology flow is shown in
Figure 1.

A. Data Collection
The first step in developing SmartWasteNet contains collecting a comprehensive and diverse set of waste images. Moreover, the
primary dataset TACO (Trash Annotations in Context) contains real-world waste images with complex backgrounds representing
multiple different categories of waste like residual, organic, recyclable, and hazardous. Additionally, to enhance generalization and
validate model performance where the dataset may be supplemented with TrashNet, that provides labeled, clean images of common
waste items. Furthermore, this ensures that the models handle both real-world and controlled urban scenarios which cover variability
in scale differences, object appearance, and background clutter.

B. Data Preprocessing
Furthermore, Preprocessing is executed to ensure that the networks learn meaningful patterns in case of environmental variability.
Moreover, Images are normalized, resized, and further data augmentation is applied by brightness adjustment, rotating, and flipping
to simulate multiple different real-world conditions. In addition, the waste categories are organized into a hierarchical structure, with
coarse-level categories such as (Organic vs Non-Organic) and fine-level categories such as (Residual, Plastic, Metal, Glass, Paper,
Hazardous). Further, this hierarchical labeling sets the foundation for the confidence-guided classification framework.
1) Normalization: Normalization is applied to standardize the pixel intensity distribution over all input images, decreasing
variations which are caused by various lighting conditions and camera sensors. Further, scaling the pixel values to a consistent
range where normalization makes numerical stability while learning and prevents certain features from dominating the training
process. Moreover, this step facilitates faster convergence and enables the model to focus on learning discriminative visual
patterns rather than being influenced by illumination inconsistencies.
2) Resizing: All input images are resized to a uniform equal spatial resolution to maintain consistency along the dataset and to
make compatibility with deep learning architecture. Moreover, Resizing allows efficient batch processing, and both local and
global features are extracted at a compatible scale. Further, this step is particularly important for allowing the model to
generalize across images containing waste items of varying dimensions and capturing size-invariant characteristics of waste
objects.
3) Data augmentation: Data augmentation is trained to improve the diversity of the training data and improve the model’s
robustness to real-world variability. moreover, Augmentation techniques like horizontal flipping, brightness adjustment, and
rotation are used to simulate different environmental conditions, object orientations, and camera viewpoints. Furthermore, this
process reduces overfitting and allows the model to learn invariant features, by improving its ability to handle viewpoint
changes, occlusion, and background clutter commonly encountered in smart city waste management scenarios.

C. Feature Extraction

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 243
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 14 Issue II Feb 2026- Available at [Link]

The preprocessed data is fed in feature extraction where a SmartWasteNet follows a hybrid feature extraction plan to efficiently
capture both local characteristics and global contextual information from waste images. Moreover, this dual-representation design
addresses the limitations of single-path feature extractors and allows for robust understanding of complex urban waste scenes.

1) Local Feature Learning using DP-CNN

The DP-CNN segment which is responsible for learning multi-scale local representations that including edge information, fine-
grained textures, and structural patterns of waste objects. Further, by processing features at multiple receptive fields DP-CNN
efficiently captures variations in object surface, size, and shape properties. Moreover, this capability overcomes the limitations of
conventional CNN architectures which often struggle to recognize detailed patterns and scale variations commonly observed in
heterogeneous and cluttered urban waste environments.

2) Global Feature Modeling using Transformer

Further, SmartWasteNet includes a Transformer-based model for global context modeling. Moreover, using self-attention
mechanisms, the transformer captures spatial relationships and long-range dependencies among the entire images. Additionally, this
allows the model to understand scene-level context and object interactions. Furthermore, these are critical for resolving ambiguities
and distinguishing similar waste categories which are caused by partial occlusion or background clutter.

D. Feature Fusion
Furthermore, the local features extracted by DP-CNN and the global contextual representations extracted by the Transformer are
conceptually fused together into a discriminative and unified feature vector. Moreover, these fusion balances both local information
with comprehensive scene understanding resulting in a robust representation that enhances downstream classification. Additionally,
by integrating completing feature perspectives SmartWasteNet gains improved reliability and generalization in real-world waste
segregation scenarios.

E. Confidence-Guided Hierarchical Classification

The fused feature vector is first classified at a coarse level into Organic or Non-Organic categories. Moreover, a confidence score is
computed for this prediction conceptually representing the model’s certainty about its decision. Further, if the confidence exceeds a
predefined threshold, then the coarse-level prediction gets finalized that reducing unnecessary computation. Furthermore, if
confidence is below the threshold, then the sample is routed to a fine-grained classification stage where it is classified into detailed
waste categories like Residual, Plastic, Metal, Glass, Paper, or Hazardous. Then, this adaptive and hierarchical mechanism
conceptually prevents misclassification, minimizes error propagation in ambiguous cases, and optimizes processing efficiency by
calculating the depth analysis to the difficulty of each sample.

F. Training and Conceptual Learning

The training and conceptual learning of the SmartWasteNet has been trained under a supervised learning framework conceptually
guided by cross entropy loss to reduce the classification errors across both coarse and fine levels. However, Early stopping or other
generalization strategies are considered to prevent overfitting. Conceptually the theory behind it is where the network learns to
jointly leverage local patterns from DP-CNN (Deep Pyramid Convolutional Neural Network) and global context from the
Transformer while using the confidence-guided hierarchy to make adaptive reliable decisions particularly in challenging urban
waste environment.

IV. EXPERIMENTAL RESULTS

The proposed SmartWasteNet framework was implemented as a software-based deep learning model to evaluate its effectiveness for
automated waste segregation in smart city environments. The system was configured with an NVIDIA GPU with CUDA support
along with a multi-core CPU for data handling and preprocessing tasks. Moreover, Adequate system memory (RAM) was allocated
to manage large-scale waste image datasets. The metrices which were used our proposed SmartWasteNet method were namely
Accuracy, Precision, Recall, and F1-score are mentioned in below equation (1) - equation (4):
Accuracy = (1)
Precision = (2)

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 244
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 14 Issue II Feb 2026- Available at [Link]

Recall = (3)
F1 – score = (4)

The above mentioned are the metrices where TP notes True Positive, TN denotes True Negative FP denotes False Positive, and FN
denotes False Negative. These metrics estimate overall correctness, balanced performance, prediction of exactness, and
classification of completeness.

A. Performance Analysis
This section presents the evaluation of proposed SmartWasteNet model Waste to evaluate its effectiveness in both extraction of
images and classification. For this, the SmartWasteNet model is evaluated against CNN-based waste classifier, Hybrid CNN
(Convolutional Neural Network) which is shown in the below table 1:

Table 1: performance Analysis of proposed SmartWasteNet method

Metrics Accuracy (%) Precision (%) Recall (%) F1-score (%)
CNN-based waste 89.6 88.9 88.2 88.5
Classifier model
Hybrid CNN model 92.1 91.3 90.9 91.1
Proposed Smart 96.1 95.3 95.0 95.1
WasteNet model

Table 1 shows the performance analysis of the proposed method SmartWasteNet with various traditional methods which are CNN-
based waste classifier, Hybrid CNN. The proposed SmartWasteNet method has achieved consistent outcomes in terms of Accuracy
(96.1%), Precision (95.3%), Recall (95.0%), and F1-Score (95.1%) respectively.

B. Comparative Analysis
The comparative Analysis section compares the performance of the proposed SmartWasteNet method against existing model
Knowledge-Based Systems of Multi-layer CNN model [11] using their respective evaluation metrices. The comparison is done to
compare the improvement over the novel method which is shown in below table 2:

Table 2: Comparative Analysis of the Proposed method

Metrics Accuracy (%) Precision (%) Recall (%) F1- Score (%)
Knowledge – Based 92.3 91.5 90.8 91.1
System Using Multi
Layer CNN model [11]
Proposed Smart Waste 96.1 95.3 95.0 95.1
Net model

Table 2 shows the comparative analysis of the proposed method against the existing Knowledge Based Systems using a Multi-layer
CNN model [11]. The exiting methods of metrics resulted in terms of Accuracy (92.3%), Precision (91.5%), Recall (90.8%), and F1-
Score (91.1%). The proposed SmartWasteNet method achieved varying results in terms of Accuracy (96.1%), Precision (95.3%),
Recall (95.0%), and F1-Score (95.1%) respectively.

V. DISCUSSION
The main goal of this research is to improve automated waste sorting in smart cities by addressing the problems of current image
classification-based waste management systems. Previous methods, including traditional CNN-based approaches reported in
Knowledge-Based Systems, relied on single-stream convolutional feature extraction and flat classification pipelines. Moreover, this
reduced the ability to face real-world challenges like changes in size, visually similar materials, messy backgrounds, and hidden
waste items. These methods also lacked the ability to make adaptive decisions and understand the global context. As a result, their

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 245
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 14 Issue II Feb 2026- Available at [Link]

classification performance was inconsistent and not very robust. To meet this goal, the proposed SmartWasteNet introduces a hybrid
deep learning framework. Moreover, this framework combines Transformer-based global context modeling with a Deep Pyramid
Convolutional Neural Network (DP-CNN) at multiple scale in extracting local features. Further, the Transformer component
captures scene-level context and long-range dependencies that help in understanding complex waste categories.
Moreover, Smart WasteNet has been assimilated with confidence-guided hierarchical classification thyeory. Here, waste images are
first sorted into broad categories, such as organic and non-organic. Furthermore, for low confidence outcomes, a classifier used for
more detailed identification including materials like plastic, metal, glass, paper, and hazardous waste. This can be modified decision
process improves robustness, increases reliability, and reduces misclassification. Finally, the experiments were performed on TACO
and TrashNet datasets where SmartWasteNet significantly vary with respect to existing methods in terms of accuracy, precision,
recall, F1-score, and resistance to occlusion while maintaining scalability for real-world application. Hence, the SmartWasteNet
framework successfully achieves the objective by offering a context-aware, adaptive, and robust solution for intelligent waste
segregation in smart city environments.

VI. CONCLUSION
This research presented SmartWasteNet which is an intelligent deep learning–based waste segregation framework fitted for smart
city environments. Moreover, the proposed model integrates Deep Pyramid Convolutional Neural Networks that to capture multi-
scale local visual features and Transformer-based global context modeling to learn long-range dependencies within complex waste
scenarios. Further, a confidence-guided hierarchical classification strategy was performed to allow adaptive coarse-to-fine waste
categorization, efficiently addressing the disadvantages of flat CNN-based pipelines, like reduced robustness under occlusion and
visual ambiguity. Experimental evaluations on benchmark waste image datasets demonstrated that the proposed framework achieves
consistent improvements in classification reliability and adaptability when compared with existing approaches. In fact, the
hierarchical architecture introduces additional conceptual complexity; it allows decision scalability and reliability for real-world
urban deployments. Moreover, future work will examine model efficiency and simplification optimization to support real-time
operations and broader smart city integration.

REFERENCES
[1] Celik, G., 2025. Multi-layer feature fusion for high-accuracy solid waste classification using a hybrid deep learning model. The Visual Computer, pp.1-23.
[2] Castro-Bello, M., Roman-Padilla, D.B., Morales-Morales, C., Campos-Francisco, W., Marmolejo-Vega, C.V., Marmolejo-Duarte, C., Evangelista-Alcocer, Y.
and Gutiérrez-Valencia, D.E., 2025. Convolutional neural network models in municipal solid waste classification: towards sustainable management.
Sustainability, 17(8), p.3523.
[3] Gude, D.K., Bandari, H., Challa, A.K.R., Tasneem, S., Tasneem, Z., Bhattacharjee, S.B., Lalit, M., Flores, M.A.L. and Goyal, N., 2024. Transforming urban
sanitation: enhancing sustainability through machine learning-driven waste processing. Sustainability, 16(17), p.7626.
[4] Chauhan, R., Shighra, S., Madkhali, H., Nguyen, L. and Prasad, M., 2023. Efficient future waste management: A learning-based approach with deep neural
networks for smart system (LADS). Applied Sciences, 13(7), p.4140.
[5] Malik, M., Sharma, S., Uddin, M., Chen, C.L., Wu, C.M., Soni, P. and Chaudhary, S., 2022. Waste classification for sustainable development using image
recognition with deep learning neural network models. Sustainability, 14(12), p.7222.
[6] Chhabra, M., Sharan, B., Elbarachi, M. and Kumar, M., 2024. Intelligent waste classification approach based on improved multi-layered convolutional
neural network. Multimedia Tools and Applications, 83(36), pp.84095-84120.
[7] Sayem, F.R., Islam, M.S.B., Naznine, M., Nashbat, M., Hasan-Zia, M., Kunju, A.K.A., Khandakar, A., Ashraf, A., Majid, M.E., Kashem, S.B.A. and
Chowdhury, M.E., 2025. Enhancing waste sorting and recycling efficiency: robust deep learning-based approach for classification and detection. Neural
Computing and Applications, 37(6), pp.4567-4583.
[8] Alkılınç, A., Okay, F.Y., Kök, İ. and Özdemir, S., 2025. Deep Ensemble Learning Model for Waste Classification Systems. Sustainability, 18(1), p.24.
[9] Dipo, M.H., Farid, F.A., Mahmud, M.S.A., Momtaz, M., Rahman, S., Uddin, J. and Karim, H.A., 2025. Real-Time Waste Detection and Classification Using
YOLOv12-Based Deep Learning Model. Digital, 5(2), p.19.
[10] Nahiduzzaman, M., Ahamed, M.F., Naznine, M., Karim, M.J., Kibria, H.B., Ayari, M.A., Khandakar, A., Ashraf, A., Ahsan, M. and Haider, J., 2025. An
automated waste classification system using deep learning techniques: Toward efficient waste recycling and environmental sustainability. Knowledge-Based
Systems, 310, p.113028.
[11] Gibellini, F., Fraternali, P., Boracchi, G., Morandini, L., Martinoli, T., Diecidue, A. and Malegori, S., 2025. A deep learning pipeline for solid waste detection in
remote sensing images. Waste Management Bulletin, p.100246.
[12] Cheema, S.M., Hannan, A. and Pires, I.M., 2022. Smart waste management and classification systems using cutting edge approach. Sustainability, 14(16),
p.10226.
[13] Wang, Z., Zhou, W. and Li, Y., 2024. GFN: a garbage classification fusion network incorporating multiple attention mechanisms. Electronics, 14(1), p.75.
[14] Nnamoko, N., Barrowclough, J. and Procter, J., 2022. Solid waste image classification using deep convolutional neural network. Infrastructures, 7(4), p.47.
[15] Naznine, M., Nahiduzzaman, M., Karim, M.J., Ahamed, M.F., Salam, A., Ayari, M.A., Khandakar, A., Ashraf, A., Ahsan, M. and Haider, J., 2025. PLDs-CNN-
ridge-ELM: Interpretable lightweight waste classification framework. Engineering Applications of Artificial Intelligence, 162, p.112522.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 246