1

Real-Time Waste Classification Using Computer

Vision
Devika Madhusoodanan Nithin Sagar Vismaya R
Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering, Engineering, Engineering,
Amrita School of Computing, Bengaluru, Amrita School of Computing, Bengaluru, Amrita School of Computing, Bengaluru,
Amrita Vishwa Vidyapeetham, India Amrita Vishwa Vidyapeetham, India Amrita Vishwa Vidyapeetham, India
devikam.280@[Link] pothulapallinithinsagar@[Link] vismaya.r88@[Link]

Hridyalakshmi Santhosh D. Radha

Department of Computer Science and Department of Computer Science and
Engineering, Engineering,
Amrita School of Computing, Bengaluru, Amrita School of Computing, Bengaluru,
Amrita Vishwa Vidyapeetham, India Amrita Vishwa Vidyapeetham, India
hridyalakshmi2004@[Link] d radha@[Link]

Abstract—Applications of deep learning methods have suc- time image recognition capabilities and automated sorting
cessfully enabled automated waste classification, improving the machinery through these systems offers major improvements
process of waste management. A dataset with biodegradable to waste recycling processes and generates reduced landfills
and non-biodegradable categories, which includes eight separate
subcategories, is used to train and test three Convolutional Neural and environmentally friendly waste management systems.
Network models: ResNet-18, VGG-16, and AlexNet. The ResNet- The use of AI systems for sorting waste enhances the speed
18 model beats the others with an accuracy of 83.76%. The of operation and environmental sustainability since it results in
models are put into use in the web interface built using Flask, increased recycling rates of valuable resources. The proposed
which provides instant image predictions based on uploaded files methodology focuses on the development of a CV-based waste
or live camera flow. The system shows the practical potential
of CNN-based classifiers to improve environmental sustainability classification system aimed at instructive identification and
through automated sorting of waste. classification of waste as biodegradable or non-biodegradable.
The system exploits pre-trained deep learning frameworks,
Index Terms—Waste Classification, ResNet-18, VGG-16,
AlexNet, Image Classification, Deep Learning, Model Evaluation, such as ResNet-18, VGG-16, and AlexNet, with a set of
Real-time Prediction, Flask Web Application images of different waste from the dataset. Using these models,
the system uses data processing and delivers instantaneous
results of classification through a Flask based web application.
I. I NTRODUCTION The users can recognize the type of waste through the ability
Waste management is a global critical concern since im- to upload pictures or access the device camera. The purpose is
proper waste disposal methods result in environmental pol- to provide a suitable solution for the management of waste for
lution in addition to resource wastage and health complica- boosting recycling and mitigating ecological harm. In addition
tions. Waste segregation methods involving human effort are to image uploads, there is real-time detection provided by the
inadequate since they are time-consuming, generate numerous tool, which means waste can be classified immediately through
errors, and are not efficient. The emergence of automation real-time detection.
and artificial intelligence technology facilitated the creation of
artificial intelligence-based waste segregation systems that en-
II. L ITERATURE S URVEY
hance waste sorting accuracy while enhancing efficiency. The
waste sorting systems incorporate computer vision in combina- Himanshu et. al presents a computer vision system for real-
tion with machine learning algorithms to divide materials into time recognition of waste printed circuit board (WPCB) elec-
four groups which are biodegradable waste, non-biodegradable tronic components (ECs) to improve e-waste recycling [1]. The
waste and recyclable waste together with hazardous waste. work reveals the environmental and economic consequences of
AI waste segregation system creates robust deep learning hazardous PCB disposal and suggests a computer vision solu-
models which learn from massive waste image databases in tion based on the YOLOv3 deep learning model. The system
order to attain high waste identification accuracy for purpose is trained on open-source PCB datasets and a newly developed
of classification. The system uses image processing techniques domestic WPCB dataset developed under real-world recycling
backed by convolutional neural networks and object detection conditions. The model shows improved classification precision
techniques in order to analyze visual aspects of shape, color and real-time detection, which minimizes the use of manual
and texture through its process. The integration of real- labor and optimizes material recovery. Further research, such
 2

as incorporating more sensors, should be pursued to further and health risks that are commonly exacerbated in developing
improve detection accuracy and sorting efficiency in recycling countries with manual modes of waste collection. While the
plants. existing studies lay emphasis on single-class waste detection,
Prashant et. al introduces a computer vision technique the work proposed in [6] takes a multi-class classification view
for real-time sorting and classification of e-waste materials using CNN, YOLO, and Faster R-CNN. The smart dustbin
employing Convolutional Neural Networks (CNN) [2]. It employing computer vision methodologies is designed using
emphasizes the economic and environmental advantages of Raspberry Pi and a camera module. From the experimental
recycling e-waste through automation by enhancing metal results, Faster R-CNN is shown to attain a maximum accuracy
extraction and plastic separation. YOLO v7 and v5 object of 91% at a loss of 16%, when compared to CNN and
detection models combined with TensorFlow are used in the YOLO. Analyzing against the existing smart dustbin models
research to yield high accuracy in the detection of copper, shows the efficacy of computer vision in enhancing the waste
printed circuit boards (PCBs), steel, glass, and aluminum. classification and recycling process.
The experiment shows a 94% accurate prediction for 240 e- Research on computer vision (CV) for municipal solid waste
waste items with optimized batch size and training options for (MSW) sorting increased dramatically by looking at smart
better performance. The study highlights the usefulness of AI- technologies like robotics for making waste management bet-
assisted e-waste classification towards more efficient recycling ter. However, the literature has limited studies regarding where
and sustainability. it came from or what the state is or what future challenges lay
Nežerka and team introduce a machine-learning-based ahead. According to a critical review of academic research in
framework for classifying fragments of construction and de- [7], the trend is that old machine learning moves into deep
molition waste (CDW) with computer vision [3]. It compares learning since that transition seems to be driven by substantial
convolutional neural networks (CNN) with some feature ex- advances in computational power and algorithms. While stud-
traction techniques applied to gradient boosting (GB) and ies create an uneven distribution across different waste sectors,
multi-layer perceptron (MLP) classifiers. The study shows that they still rely too much on reasonably simplified environments
feature extraction greatly enhances classification speed and and artificially collected data. Future research should rather
accuracy, beating CNN models, especially for visually similar consider such real-world complexities and possibilities to open
materials such as concrete and autoclaved aerated concrete for collective sharing of waste image datasets for greater CV
(AAC). GB classifier recorded the maximum accuracy of application in waste sorting at industrial settings.
92.3%, while CNN trailed behind at 85.9%. The results The literature on waste management emphasizes the prob-
demonstrate the potential of CDW sorting with AI-powered lems that doped waste segregation creates, resulting in added
systems for enhanced recycling and sustainability. environmental and health concerns, particularly in vulnerable
Thangam et. al shows a computer vision-based waste gath- populations. In Honduras, only 2% of waste is recycled due to
ering and segregation platform for ships via a Convolutional horrible conditions in the informal sector. In [8], we propose a
Neural Network (CNN) [4]. It seeks to solve water pollu- computer vision system to perform real-time classification of
tion through the automatization of refuse collection on sun- dry wastes, targeting aluminum cans, PET bottles, and glass
powered boats. The system is based on using a conveyor bottles. The use of a dataset of 5,981 images generated through
belt system to sweep waste from bodies of water and then data augmentation earned the system a mean average precision
sorting out the biodegradable and non-biodegradable waste via (mAP) of 0.95, with accuracy and recall values above 85%.
CNN-based image recognition. The model proposed attained The model encourages fast inference at 16.5 ms per image,
a classification accuracy of 90.03%, with room for improve- thus paving the way for automated waste segregation.
ment through dataset expansion. The research emphasizes the Jin et. al developed a MobileNetV2 model enhancement
system’s capacity to minimize human involvement in waste [9] which includes the convolutional attention mechanism
management and promote sustainable marine environments. (CBAM) alongside principal component analysis (PCA) and
Now Researchers come together under a cross review transfer learning for better generalization. The developed
framework to clean construction material datasets in concrete, model demonstrates 90.7% accuracy on ”Huawei Cloud”
brick, metal, wood, and stone for effective data quality in dataset while running at 600 ms on a Raspberry Pi 4B and
machine learning. By collecting and classifying images into requiring 30.1% less storage than MobileNetV2. The real-
main material, background, and uncertainty categories, they world testing of developed garbage sorting system produced
achieve the reliability of the dataset. [5] evaluates the dataset a 89.26% accuracy rate. The approach creates its innovative
quality under a number of deep convolutional neural networks power by combining CBAM with PCA and transfer learning
such as VGG16, GoogleNet, and ResNet in mobile friendly capabilities that simultaneously increase precision and de-
and unfriendly architectures. Experimental results confirm the crease system requirements and let real-time function on edge
effectiveness of datasets and demonstrate the necessity of data equipment. The study shows how lightweight deep learning
cleaning before the training of deep learning models. Future conducts waste recycling tasks efficiently while developers
studies call for the integration of the deep construction material aim to add Yolov5 for multiple detection functionality together
recognition model with the cross-review framework to perform with system optimization improvements.
better. Mao et al. developed a waste classification system through
Recycling research that deals with waste classification gives optimized DenseNet121 model implementation that uses
importance to segregation so as to prevent cross-contamination CNNs for automated waste sorting applications [10]. Data aug-
 3

mentation strategies involve flipping trash items horizontally the study identifies 15 key enablers validated through expert
and vertically while rotating them randomly for the Trash- opinions, analyzing their driving-dependence powers (MIMAC
Net dataset expansion process. An optimized DenseNet121 analysis) and cause-effect relationships (DEMATEL analysis).
model receives its fully-connected layers adjustment through a The findings suggest that CVT-driven automation can improve
genetic algorithm (GA) optimization of neuron numbers and classification, sorting, and disposal, reducing manual labor
dropout values. The optimized DenseNet121 model reaches costs, enhancing safety, and ensuring real-time monitoring.
99.6% accuracy performance while designing the best CNNs Additionally, CVT integration supports circular economy prin-
available and uses Grad-CAM to detect essential waste image ciples, optimizing material recovery and minimizing environ-
characteristics. PLASRec achieves better classification out- mental impact. Policymakers should implement centralized
comes because it integrates GA-based hyperparameter op- databases and intelligent decision-making systems to enhance
timization and restored fully-connected layers in a novel process control and efficiency. Ultimately, adopting CVT in
manner. Future developments of this waste sorting method e-waste management offers economic, environmental, and
will address weaknesses in prediction accuracy and improve operational benefits, paving the way for sustainable waste
the robustness of the collected dataset. processing solutions.
Wang et. al presents a MWM system that combines deep Zhiming Dong and his team proposed Boundary-Aware
learning together with cloud computing and IoT tools to Transformer (BAT) framework enhances semantic segmenta-
enhance waste sorting and monitoring as well as collection tion for construction waste by integrating morphological oper-
systems [11]. The system achieves 94.24% accuracy through ations, a Transformer-based segmentation model, and a deep
CNN-based processing with MobileNetV3 as its top per- learning-based boundary refinement scheme [14]. It outper-
former, because it operates at 261.7 ms and maintains a forms the DeepLab v3+ baseline by 5.48% in MIoU, demon-
compact model size of 49.5 MB. Connected garbage bins strating superior feature extraction and boundary processing
containing IoT sensors track waste quantity and toxic emis- capabilities. Despite these improvements, challenges like cat-
sion data which the cloud platform analyzes to make instant egory imbalance and segmentation errors persist. Future work
operational decisions. The primary innovation of this system can focus on data augmentation, improved annotation quality,
stems from its combination of advanced CNNs for waste and advanced model architectures. The study highlights BAT’s
classification precision with IoT monitoring capabilities and effectiveness in fine-grained material recognition, making it
cloud-processing data in an efficient manner. The system valuable for practical waste management applications.
produces accurate results while demonstrating efficient per- Thangam and her team propsed implementation of an
formance which indicates its potential deployment in real- Arduino Uno-based metallic waste management system has
life applications. The authors plan to study how detection of demonstrated significant improvements in waste segregation,
multiple waste objects can work within one image and develop collection, and recycling efficiency [15]. Real-time sensor
motivational strategies that encourage people to help with the data enables precise waste identification, optimizing sorting
waste sorting process. mechanisms and reducing manual intervention. Automated
Sasikanth et. al presents an efficient smart waste manage- transportation and centralized monitoring enhance logistics,
ment solution which automates household waste sorting to minimizing environmental impact. The integration of IoT-
handle environmental and health issues related to improper based notifications ensures timely waste disposal, preventing
waste management practices [12]. An Arduino microcontroller overflow and inefficiencies. Future enhancements, such as con-
together with Raspberry Pi operates as the basis of waste veyor belts, moisture sensors, and renewable energy sources,
classification into wet dry and metal waste while additional can further refine waste management practices. Overall, the
image processing separates dry waste into paper plastic and system contributes to sustainability by promoting resource
glass categories. The system uses moisture and metal sensor recovery and reducing dependence on raw materials.
devices together with TensorFlow-based algorithms SSD and Wanqi Ma and his team proposed DSYOLO-Trash model
Faster-RCNN to attain precise identifications. A waste data which significantly improves solid waste detection accuracy,
tracking platform operates within an online interface that achieving higher mAP and recall rates than YOLOv5 vari-
requires users to pay fees based on their waste disposal ants [16]. Integrating CotNet and CBAM enhances feature
activities to promote proper waste management. The system extraction, enabling better detection of small, overlapping,
achieved high detection accuracy where Faster-RCNN outper- and occluded objects. Experimental results confirm its robust-
formed precision while SSD proved optimal for Raspberry Pi ness in complex backgrounds and multi-object scenarios. The
because of its reduced memory requirements. Real-time waste model, combined with DeepSORT tracking, ensures efficient
classification and a user-friendly web interface combined with real-time waste identification. A smart detection and sorting
hardware and software form the distinctive features of this system using robotic arms demonstrates practical applicability
system which provides households with an effective solution in waste management. Future work can optimize speed and
for waste segregation and promotes recycling. efficiency for broader real-world deployment.
Himanshu Sharma and peers explores the role of Computer
Vision Technology (CVT) in enhancing e-waste management
III. M ETHODOLOGY
by addressing challenges such as ineffective regulations, lack
of awareness, and unsafe handling practices in developing The method and procedures that were employed to construct
nations [13]. Using an ISM-DEMATEL-based framework, and test a deep learning model for classifying waste is outlined
 4

in this section. A wide variety of methodology has been 3) AlexNet: One thing that makes it stand out is its contri-
implemented, including the way the dataset is organized, pre- butions, albeit its age and shallower structure, computational
processing measures, the choice of a model, the application benefits, and integration of features such as ReLU, dropout,
of training measures, evaluation and the deployment of the and overlapping pools.
system for real-time predictions. Transfer learning was applied to each model by leaving
the initial layers unchanged and adding the final classification
A. Dataset Preparation layer (adjustable to the target number of classes). By this
The dataset used in this research was structured in a approach, it is possible for the models to include previous
hierarchical form. The images were divided into two cate- knowledge of identifying basic image pattern so as to adjust
gories, in which the collection was made up of two classes: its forecasts towards classification needs of the target domain.
biodegradable and non-biodegradable, each containing multi-
ple subcategories such as food waste, paper, plastic and metal.
D. Model Training Strategy
Two major parts of the dataset were: train and val which are,
respectively, the training and validation dataset. Training the models meant making use of the supervised
By structuring images in subfolders according to class and learning techniques, which is applied to the training dataset.
subcategory, the system automatically tagged images which Batches of images were passed through neural networks during
made it easier to train convolutional neural networks that could training, and a loss function was applied against the true labels
identify several waste types. The class label of each image was for the measurement of the neural network’s predictions.
given based on the subcategory existing in its directory. 1) Loss Function: When faced with the job of predicting
class probabilities, cross-entropy loss was an appropriate one
B. Data Preprocessing and Augmentation to minimize the difference between predicted and actual values
To ensure that there is uniform input and strengthen the within several classes.
ability of the model to generalize, a series of preprocessing 2) Optimizer: An approach to update model weights that
techniques were followed: were adjusted according to calculated gradients was used to
1) Resizing: The sizes of images were uniformized based train the model. This approach enhanced convergence rate
on a predetermined size that the neural networks could handle without reducing the overall stability of the training.
effectively. 3) Learning Rate and Epochs: The models finished training
2) Normalization: The standardization of pixel intensity in after a predetermined epoch count, thus guaranteeing learning
all images also allowed for faster training and maintained rate calibration. A portion of effort was made to prevent
model’s numerical accuracy. overfitting and encourage efficient training through learning
3) Augmentation: During the training process many aug- rate scheduling policy and early stopping strategy.
mentation strategies were used such as horizontal flipping of On every training iteration, the sizes of the validation dataset
images, random rotations, adjusting the brightness. Employing were measured and carried out to assess the generalization of
these strategies mirrors actual world variations in orientation, each model. Even without GPU support in light of hardware
scale, and lighting, which, among other things, provides a controls, the training had been conducted within an extended,
solution to overfitting and increases the model’s performance but not entirely unsatisfactory, period.
in new data.
By adhering to the preprocessing pipeline, the models
E. Model Evaluation
received sanitized, standardized and cleansed data, augmented
with synthetic variations. Post training, performance metrics were used on the vali-
dation dataset to evaluate the models.
C. Model Selection and Transfer Learning 1) Accuracy: It is the ratio of correctly classified images
Three widely known convolutional neural network struc- to an overall number of images in a validation set.
tures, namely ResNet-18, VGG-16, and AlexNet, were used TP + TN
in the proposed methodology and their performance were Accuracy = (1)
TP + TN + FP + FN
evaluated. Before being tuned for the waste classification
task, these architectures were trained on ImageNet dataset. It served as a rough measure of model performance, provid-
The models were selected according to their differentiated ing an approximation of the frequency with which classes
complexity, depth, and computational resources needed. predicted were reflected in reality. Although obtaining high
1) ResNet-18: It utilizes residual connections that make accuracy is important, it frequently does not reflect a model’s
smooth flow of gradients during the back propagation process true performance, which can happen often when the variety of
possible. Utilizing these connections, the model can be trained classes in a dataset is unevenly distributed.
deeply without loss in performance. In order to obtain a deeper insight into the performance of
2) VGG-16: It consists of a simple structure, with various models, three additional metrics were added.
and identical con-volutional layers all using small filters. 2) Precision: It counted the percentage of correct positive
The deep representation depth of VGG-16 and its ability to classifications versus total instances assigned to a given class.
effectively model complex hierarchical visual features have TP
been demonstrated. Precision = (2)
TP + FP
 5

High precision meant that the model never or rarely misiden- a reference file. With this strategy, the system automatically
tified images, and therefore sorted correctly in the case where chose the best model for real-time inference and was no longer
errors caused waste to go into the wrong container. required to be manually monitored.
3) Recall: It measured how well the model could predict
and categorize correctly each occurrence of any given class. G. Real-Time Deployment
TP When a proper model had been selected based on eval-
Recall = (3)
TP + FN uation, it was integrated into a real-time waste classifying
High recall score meant that the model identified a consider- platform that runs on a flask web application. Two central
able number if not all the images related to a certain class thus functionalities in the system were designed:
minimising the number of false negatives. 1) Image Upload: The system allowed users to upload
4) F1-Score: It weighted-averaged the performance of pre- pictures of waste and after that the pictures were prepped and
cision and recall. labeled through the trained model.
2) Camera Interface: Users could view live video from a
2 × Precision × Recall
F1-Score = (4) webcam in the application, choose waste objects, and receive
Precision + Recall real-time classification results within the application too.
Selection of this metric proved to be beneficial in instances The system was developed so that it uses a model with
when the moderate balance of precision and recall achieved the highest evaluation results in real-time processing. After
was important, especially in cases when twice-distributed an image had been fed in, the model performed inference and
classes were used. A high F1 – score indicated that the model reported the category of waste predicted – that is biodegradable
performed well in terms of precision and recall rate for every or non-biodegradable, also the finer subcategory.
category of waste.
The assessment of these performance measures was done IV. R ESULTS
separately for every waste class – biodegradable and non- Three CNN architectures were used to assess the suggested
biodegradable. This allowed for a comprehensive evaluation AI trash classification system: AlexNet, VGG-16, and ResNet-
of model performance in all waste classifications. 18. Metrics like accuracy, precision, recall, F1-score, and
5) Confusion Matrix: In order to determine how well confusion matrix were taken into account in this assessment.
classification was done, a confusion matrix was created and Throughout the training procedure, all three models’ training
analyzed for each model. The matrix assured the figures of and validation accuracies were continuously observed. ResNet-
actual and allegedly estimated classes were arranged in a 18 stabilised at 83.76%, the highest validation accuracy among
table representing where each row represented actual class and them. According to Table I, VGG-16 achieved a peak val-
each column represented a predicted class. Diagonal entries idation accuracy of 56.87%, whereas AlexNet recorded a
indicated correct classifications and those outside the diagonal relatively lower accuracy of 19.07%.
wrongly classified. An example of this may be illustrated when
the system does not differentiate between plastic waste and the TABLE I: Performance Comparison of Waste Classification
metal causing a misclassification reflected in the corresponding Models
matrix cell. These patterns exposed the regular confusions
Model Accuracy
amongst specific class pairs, which helped identify errors ResNet-18 84.00%
that might guide future data correction or model amendment VGG16 56.87%
processes. AlexNet 19.07%
6) Evaluation Summary and Best Model Selection: During
evaluation, phase models could be evaluated and ranked on The most dependable model for this classification test
their outputs in validation. Only accuracy gave a general pic- was ResNet-18, which demonstrated superior generalisation
ture, but the precision, recall, F1-score and confusion matrix on the validation set with little indications of overfitting.
together revealed deeper and more complete proofs on the In comparison to VGG-16 and AlexNet, it demonstrated a
trustworthiness and reliability of the model in each category. smaller final validation loss and faster convergence. AlexNet
Model that had the better performance on all balanced demonstrated the largest final validation loss, suggesting weak
metrics was selected as the best candidate to be implemented. generalisation power, while VGG-16 showed somewhat higher
After its selection, the highest performing model received the loss fluctuations. ResNet-18’s validation loss, on the other
title “best model” and was being saved to be put into use hand, converged to a low and stable value. The tables II, III,
immediately, and it was used to classify tasks in real time and and IV provide the specific performance metrics for ResNet-
end application development. 18, VGG16, and AlexNet on the trash classification job.
All three models’ confusion matrices were produced, and
the results showed that ResNet-18 fared better than the others
F. Best Model Selection in accurately categorising every category with fewer misclassi-
An automated evaluation script was developed to evaluate fications. Across all classes, it showed excellent precision, but
and rank all models trained validation accuracy, enabling au- it excelled at differentiating between biodegradable (like food
tomatic identification of the best. The top validation accuracy waste in paper) and non-biodegradable (like plastic, metal)
model was chosen automatically and its name was entered into things. ResNet-18 successfully reduced false negatives in the
 6

majority of categories while achieving high recall values as

well. Its resilience in managing inter-class imbalance was
further demonstrated by the balanced F1-scores for every
class. Because of their visual similarities, plastic and metal
trash were frequently misclassified in VGG-16 and AlexNet.
ResNet-18’s stronger feature extraction and generalisation
skills were highlighted by the considerable reduction in these
errors.
TABLE II: Classification Report for ResNet-18
Class Precision Recall F1-Score Support
biodegradable/food waste 0.73 0.96 0.82 229
biodegradable/leaf waste 0.97 0.97 0.97 394
biodegradable/paper waste 0.92 0.72 0.81 212
biodegradable/wood waste 0.81 0.64 0.72 59
non biodegradable/ewaste 0.93 0.69 0.79 55
non biodegradable/metal cans 0.78 0.74 0.76 69
non biodegradable/plastic bags 0.86 0.45 0.59 53
non biodegradable/plastic bottles 0.65 0.78 0.71 130
Accuracy 0.84 1201
Macro avg 0.83 0.74 0.77 1201
Weighted avg 0.85 0.84 0.84 1201
Fig. 1: Prediction of food waste image in the flask web-app

TABLE III: Classification Report for VGG16

Class Precision Recall F1-Score Support
biodegradable/food waste 0.48 0.96 0.64 229
biodegradable/leaf waste 0.93 0.84 0.88 394
biodegradable/paper waste 0.66 0.34 0.45 212
biodegradable/wood waste 0.00 0.00 0.00 59
non biodegradable/ewaste 0.56 0.16 0.25 55
non biodegradable/metal cans 0.21 0.71 0.32 69
non biodegradable/plastic bags 0.12 0.06 0.08 53
non biodegradable/plastic bottles 0.00 0.00 0.00 130
Accuracy 56.87% 1201
Macro avg 0.37 0.36 0.38 1201
Weighted avg 0.38 0.57 0.33 1201

TABLE IV: Classification Report for ALEXNET

Class Precision Recall F1-Score Support
biodegradable/food waste 0.19 1.00 0.32 229
biodegradable/leaf waste 0.00 0.00 0.00 394
biodegradable/paper waste 0.00 0.00 0.00 212
biodegradable/wood waste 0.00 0.00 0.00 59
non biodegradable/ewaste 0.00 0.00 0.00 55
non biodegradable/metal cans 0.00 0.00 0.00 69
non biodegradable/plastic bags 0.00 0.00 0.00 53
non biodegradable/plastic bottles 0.00 0.00 0.00 130
Accuracy 19.07% 1201
Fig. 2: Prediction of plastic bottle waste via image capture in
Macro avg 0.02 0.12 0.04 1201 the web-app
Weighted avg 0.04 0.19 0.06 1201

Based on comprehensive evaluation metrics and validation As seen in Fig. 3, the real-time detection shows that
accuracy, ResNet-18 was chosen as the final model to be the system can handle streaming data and correctly identify
deployed. For live inference on uploaded photos and webcam objects from video frames. Camera-captured results preserve
streams, it was incorporated into the Flask-based application. the same two-stage classification methodology as Fig. 2,
Effective and automated garbage sorting was made possible by confirming consistent performance with photographed items.
the model’s accurate classification of waste into biodegradable The entire hierarchical output is shown in each result, with the
and non-biodegradable categories as well as into subcategories. principal biodegradability category and the particular material
The example outputs show how the Flask application uses classification coming next.
a hierarchical pipeline to process waste classification. The By using unified preprocessing and model inference
algorithm first assesses the biodegradability of each given pipelines, the backend design effectively manages these vari-
image, whether it was uploaded, taken with a live camera, ous input methods. The live detection examples demonstrate
or photographed, before determining the precise waste kind. that results are returned quickly enough to accommodate
The food waste image example in Fig. 1 correctly navigates real-time use cases. A unified user experience is ensured
to the appropriate food waste category via the biodegradable by the interface’s uniformity across all input types (uploads,
branch. camera images, and live detection). The system’s capacity
 7

R EFERENCES
[1] Himanshu Sharma, Harish Kumar, A computer vision-based system for
real-time component identification from waste printed circuit boards,
Journal of Environmental Management, Volume 351,2024, 119779,
ISSN 0301-4797, [Link]
[2] Prashant K. Sarswat, Rahulkumar Sunil Singh, Subbu Venkata Satyasri
Harsha Pathapati, Real time electronic-waste classification algorithms
using the computer vision based on Convolutional Neural Network
(CNN): Enhanced environmental incentives, Resources, Conserva-
tion and Recycling, Volume 207, 2024, 107651, ISSN 0921-3449,
[Link]
[3] V. Nežerka, T. Zbı́ral, J. Trejbal, Machine-learning-assisted classification
of construction and demolition waste fragments using computer vision:
Convolution versus extraction of selected features, Expert Systems with
Applications, Volume 238, Part A, 2024, 121568, ISSN 0957-4174,
[Link]
[4] Thangam, S., Gurupriya, M., Yadav, C.S.K., Harshith, M. and Ku-
mar, M.S., 2024, October. A Smart Waste Management System Using
ESP8266. In 2024 4th International Conference on Sustainable Expert
Fig. 3: Real time predicition of Plastic Bottle Systems (ICSES) (pp. 1-5). IEEE.
[5] Ying Sun, Zhaolin Gu, Using computer vision to recognize construc-
tion material: A Trustworthy Dataset Perspective, Resources, Conser-
vation and Recycling, Volume 183, 2022, 106362, ISSN 0921-3449,
to direct appropriate waste disposal through simple visual [Link]
[6] Narayanswamy, Neeraja, A. Rajak, and Shazia Hasan. ”Development of
communication is illustrated by example outputs. computer vision algorithms for multi-class waste segregation and their
analysis.” Emerging Science Journal 6.3 (2022): 631-646.
[7] Weisheng Lu, Junjie Chen, Computer vision for solid waste
V. C ONCLUSION sorting: A critical review of academic research, Waste Man-
An AI-powered waste categorization system was success- agement, Volume 142, 2022, Pages 29-43, ISSN 0956-053X,
[Link]
fully implemented and tested using three recent convolu- [8] F. Núñez, ”Dry Waste Segregation using Computer Vision,”
tional neural network (CNN) models: ResNet-18, VGG-16, 2022 IEEE Central America and Panama Student Conference
and AlexNet. Based on an extensive performance comparison (CONESCAPAN), San Salvador, El Salvador, 2022, pp. 1-6, doi:
10.1109/CONESCAPAN56456.2022.9959238.
using accuracy, precision, recall, F1-score, and confusion [9] Jin, S., Yang, Z., Królczykg, G., Liu, X., Gardoni, P. and Li, Z., 2023.
matrices, ResNet-18 showed the best overall performance. It Garbage detection and classification using a new deep learning-based
reported the highest validation accuracy of 83.76%, had lesser machine vision system as a tool for sustainable waste recycling. Waste
Management, 162, pp.123-130.
validation loss, and provided balanced class-wise F1-scores. [10] Mao, W.L., Chen, W.C., Wang, C.T. and Lin, Y.H., 2021. Recycling
These findings suggest that ResNet-18 has better feature waste classification using optimized convolutional neural network. Re-
extraction and generalization performance and is hence better sources, Conservation and Recycling, 164, p.105132.
[11] Wang, C., Qin, J., Qu, C., Ran, X., Liu, C. and Chen, B., 2021. A
suited for practical deployment. Thus, it was incorporated smart municipal waste management system based on deep-learning and
into a real-time application based on Flask that could classify Internet of Things. Waste Management, 135, pp.20-29.
waste from both uploaded images as well as real-time webcam [12] Sasikanth, B.S., Yoshita, L.N., Reddy, G.N. and Manitha, P.V., 2021,
November. An Efficient & Smart Waste Management System. In 2021
streams. The implemented system was successful in differen- International Conference on Computational Intelligence and Computing
tiating between biodegradable and non-biodegradable waste Applications (ICCICA) (pp. 1-6). IEEE.
and classifying waste further into respective subcategories, [13] Sharma, Himanshu, Harish Kumar, and Sachin Kumar Mangla. ”En-
ablers to computer vision technology for sustainable E-waste manage-
thus enabling effective and automated sorting of waste. This ment.” Journal of Cleaner Production 412 (2023): 137396.
solution takes a major leap towards encouraging smart and [14] Dong, Zhiming, Junjie Chen, and Weisheng Lu. ”Computer vision to
sustainable waste management practices with the help of AI. recognize construction waste compositions: A novel boundary-aware
transformer (BAT) model.” Journal of environmental management 305
(2022): 114405.
VI. F UTURE W ORK [15] Thangam, I. S. S. Reddy, K. B. Prakash and P. H. Reddy, ”A
Holistic Approach to Handling Metallic Waste for Environmen-
Increasing the dataset with additional diverse types of waste tal Sustainability,” 2024 3rd International Conference for Advance-
and actual environment conditions will increase the robustness ment in Technology (ICONAT), GOA, India, 2024, pp. 1-6, doi:
10.1109/ICONAT61936.2024.10774699.
of the model. Adding a multi-language user interface can [16] Ma, Wanqi, et al. ”DSYOLO-trash: An attention mechanism-integrated
make it more accessible to users in regions where other and object tracking algorithm for solid waste detection.” Waste Man-
languages dominate. Furthermore, optimizing the model for agement 178 (2024): 46-56.
edge deployment on low-power devices like Raspberry Pi can
make offline and portable use cases possible. The system
could be integrated with physical devices like robotic arms
or conveyor systems to enable end-to-end automated sorting
of waste. An active learning user feedback loop could also
be added to enhance the accuracy of classification over time.
Lastly, creating a sustainability dashboard to track waste
categorization trends would offer actionable insights and aid
in larger environmental initiatives.