SUSTAINABLE COOKING WITH AI:

GENERATIVE RECIPES AND SMART

NUTRITION ASSISTANCE

SEMINAR REPORT

Submitted By

NEEHAR CHANDRAN A(TKR22CS064)

to

the APJ Abdul Kalam Technological University

in partial fulfillment of the requirements for the award of the Degree

of

Bachelor of Technology in
Computer Science & Engineering

Department of Computer Science & Engineering

College of Engineering Trikaripur
Cheemeni.P.O, Kasaragod Dist., Kerala-671313.

July, 2025

An ISO 9001-2015 Certified Institution Accredited by NAAC,NBA

 DECLARATION

I hereby declare that the seminar report for ”Sustainable Cooking With AI: Gen-
erative Recipes And Smart Nutrition Assitance” submitted for partial fulfillment of
the requirements for the award of the degree of Bachelor of Technology from the
APJ Abdul Kalam Technological University, Kerala, is a bona fide work carried out
by me. This submission represents my ideas in my own words, and where ideas
or words of others have been included, I have adequately and accurately cited and
referenced the original sources. I also declare that it adheres to the ethics of aca-
demic honesty and integrity and that I have not misrepresented or fabricated any
data, idea, fact, or source in my submission. I understand that any violation of
the above will result in disciplinary action by the institute and/or the University
and may also lead to penal action from the sources that have not been properly
cited or from whom proper permission has not been obtained. This report has not
previously formed the basis for the award of any degree, diploma, or similar title
from any other University.

Place: Cheemeni NEEHAR CHANDRAN A

Date: 03-10-2025

i
 DEPARTMENT OF COMPUTER SCIENCE &
ENGINEERING
COLLEGE OF ENGINEERING TRIKARIPUR

CERTIFICATE

This is to certify that the report entitled SUSTAINABLE COOKING

WITH AI: GENERATIVE RECIPES AND SMART NU-
TRITION ASSISTANCE submitted by NEEHAR CHANDRAN
A(TKR22CS064) to the APJ Abdul Kalam Technological University in partial
fulfillment of the requirements for the award of the Degree of Bachelor of Technology
in Computer Science & Engineering is a bonafide record of the project work carried
out by him/her under my/our guidance and supervision. This report in any form
has not been submitted to any other University or Institute for any purpose.

Seminar Guide Coordinator Head of the Department

Prof. Rafeek A P Prof. Rafeek A P Dr. Naveena A K

Assistant Professor Assistant Professor Associate Professor

ii
 ACKNOWLEDGEMENT

We sincerely thank the ALMIGHTY for his shower of blessings, which has enabled
us to attain success in this project.

We are extremely grateful to Dr. Mahesh V V, Principal, College of Engineering

Trikaripur, Cheemeni, providing the necessary facilities.

With immense pleasure and heartiest gratitude, we express our sincere thanks to
Dr. Naveena A K, Associate Professor and Head of CSE Department.

We are pleased to show our indebtedness to our Seminar co-ordinators [Link]

A P, Assistant Professor in CSE & Prof. Huda Noor dean, Assistant Professor
in CSE, and guide Prof. Rafeek A P, Assistant Professor in CSE Department for
his valuable suggestions and proper guidance.

I would like to thank my Class Tutor,[Link] S , for his exceptional guid-

ance, constructive criticism, and unwavering encouragement throughout this jour-
ney. I also appreciate the support from the faculty members, as well as the technical
and administrative staff of the Department of Computer Science and Engineering,
whose cooperation has been invaluable.

I am immensely thankful to my family and friends for their mental support and
motivation, which enabled me to work efficiently. Lastly, I wish to acknowledge my
parents for their unwavering love, guidance, and moral support, which has been the
foundation of my success. I express my gratitude to all other staff members and
classmates for their assistance and encouragement.

NEEHAR CHANDRAN A

iii
 VISION
(Institution)

To be a premier institution in education and research for moulding technically com-

petent and socially committed professionals.

MISSION
(Institution)

(i) Promote interdisciplinary research and innovation so as to meet the current

needs of industry and society.

(ii) Attract, nurture and retain the best faculty and technical manpower.

(iii) Provide state of art facility for quality technical education.

(iv) Develop personality and professional skills of the students through interaction
with alumni academia and industry.

iv
 VISION
(Department)

To mould technically competent and socially committed professionals in the field of

computer science.

MISSION
(Department)

(i) To provide a strong foundation in theoretical and practical aspects of computer

science.

(ii) To impart technical skills necessary to generate quality professional according

to industry needs.

(iii) To develop human resource with the ability to apply the knowledge for the
benefit of the society.

v
 ABSTRACT

Cooking sustainably is key to eco-friendly living and reducing food waste. This
paper introduces an AI powered platform that combines food classification, recipe
generation and nutrition assistance for sustainable cooking. The platform uses a
convolutional neural network (CNN) for Indian food classification, RoboFlow API
for international food detection and a GPT based natural language processing (NLP)
model for personalized recipe generation. It also has waste tracking and meal plan-
ning features to optimize ingredient usage and reduce waste. The system achieved
92 percent accuracy for Indian food items and was able to detect international
food through RoboFlow. The platform generates personalized recipes based on user
inputs and available ingredients for both sustainability and healthy [Link] inter-
active Streamlit app is the user interface for food recognition, recipe generation and
sustainability focused meal planning. This system combines AI, APIs and sustain-
ability focused innovations to be a one stop solution for modern eco conscious users.

Keywords- Sustainable Cooking, Artificial Intelligence, Food Classification, Gen-

erative Recipes, Nutrition Assistance, Waste Management. .

vi
 CONTENTS

1 INTRODUCTION 1

2 LITERATURE SURVEY 2

3 METHODOLOGY AND PROPOSED SYSTEM 3

3.1 CNN (Convolutional Neural Network) . . . . . . . . . . . . . . . . . . 3
3.1.1 Input Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.1.2 Convolutional Layer . . . . . . . . . . . . . . . . . . . . . . . 4
3.1.3 Activation Function (ReLU) . . . . . . . . . . . . . . . . . . . 5
3.1.4 Pooling Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1.5 Fully Connected Layers . . . . . . . . . . . . . . . . . . . . . . 7
3.2 NLP (Natural Language Processing) . . . . . . . . . . . . . . . . . . 8
3.3 GPT-based Language Model (Generative Recipe Generation) . . . . . 9
3.4 RoboFlow Pre-trained Image Detection API . . . . . . . . . . . . . . 9
3.5 Softmax Classification . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.6 PROPOSED SYSTEM . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.6.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.6.2 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . 11
3.6.3 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . 12
3.6.4 Data Labelling . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.6.5 Training and Testing . . . . . . . . . . . . . . . . . . . . . . . 14
3.6.6 Personalised Recipe Generation . . . . . . . . . . . . . . . . . 15
3.6.7 System Framework . . . . . . . . . . . . . . . . . . . . . . . . 17
3.6.8 System Deployment and User Interface Design . . . . . . . . . 18

4 RESULTS AND DISCUSSION 19

5 ADVANTAGES AND DISADVANTAGES 22

5.1 ADVANTAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.2 DISADVANTAGES . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

vii
6 FUTURE SCOPE 24

7 CONCLUSION 25

viii
 LIST OF FIGURES

3.1 CNN-Convolutional Layer. . . . . . . . . . . . . . . . . . . . . . . . . 5

3.2 CNN-ReLU Activation. . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.3 CNN-Pooling Layer. . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.4 CNN-Fully Connected Layer . . . . . . . . . . . . . . . . . . . . . . . 8
3.5 Functioning of natural language processing . . . . . . . . . . . . . . . 16
3.6 Architectural system . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

ix
 ABBREVIATIONS
• AI - Artificial Intelligence

• CNN - Convolutional Neural Network

• NLP - Natural Language Processing

• GPT - Generative Pre-trained Transformer

• API - Application Programming Interface

• FAO - Food and Agriculture Organization (of the United Nations)

• LLM - Large Language Model

• EHFR Net - Efficient Hybrid Food Recognition Net

x
Seminar Report 2025 Sustainable Cooking with AI

CHAPTER 1
INTRODUCTION

Food waste is one of the most critical global challenges, with profound social,
economic and environmental consequences. According to the Food and Agriculture
Organization (FAO), almost one-third of all food produced worldwide is wasted, ap-
proximately 1.3 billion tonnes per year. This wastage occurs across the entire food
supply chain, from production and distribution to consumption. At the household
level, poor meal planning, excessive purchasing, and mismanagement of ingredients
are among the leading causes. Beyond the loss of edible food, this also results in
the wastage of vital resources such as water, energy, and labor involved in food pro-
duction. Furthermore, food discarded in landfills generates methane gas, a potent
greenhouse gas that significantly contributes to climate change.

As the global population continues to grow, the demand for food is increasing,
making it essential to adopt sustainable practices that reduce waste and ensure food
security. Traditional approaches, such as awareness campaigns and manual waste
tracking, have proven insufficient to address this complex issue on a large scale.
This has created the need for innovative technology-driven solutions that can make
sustainable cooking practical and accessible to people and communities.

Artificial intelligence (AI) has emerged as a transformative tool in solving sustain-

ability challenges. In the context of cooking, AI can analyze large datasets, recognize
food items, generate personalized recipes, and provide nutritional and environmen-
tal insights in real time. By integrating such features into everyday cooking, AI
empowers users to make smarter and more sustainable choices.

This report introduces an AI-powered sustainable cooking platform that combines

food recognition, personalized recipe generation, meal planning, and waste track-
ing in a single application. Designed with a user-friendly interface, the system
encourages eco-friendly cooking practices while promoting health and reducing en-
vironmental impact. By bridging the gap between technology and sustainability,
this platform represents a forward-thinking solution to one of the most pressing
problems in the world, food waste.

Department of CSE 1 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

CHAPTER 2
LITERATURE SURVEY

The issue of food waste has inspired researchers to develop advanced AI-
driven solutions to promote sustainability in food consumption and cooking. Early
research in food image recognition relied on traditional computer vision techniques,
but a breakthrough came with AlexNet by Krizhevsky et al. (2012), which in-
troduced Convolutional Neural Networks (CNNs) for image classification. Later,
ResNet by He et al. (2016) improved model depth and accuracy, leading to more
robust food recognition systems. Bossard et al. (2014) developed the Food-101
dataset, consisting of 101 food categories, which became a benchmark for food recog-
nition research. These advancements enabled real-time identification of ingredients
and dishes, which is crucial for sustainable cooking applications.

In the field of recipe understanding and generation, Salvador et al. (2017) introduced
Recipe1M, a large-scale dataset linking food images with structured recipes. This
work enabled cross-modal retrieval, where an image of a dish can be matched to its
recipe. With the rise of transformer-based models, Vaswani et al. (2017) proposed
the Transformer architecture, which became foundational for Natural Language Pro-
cessing (NLP) tasks.

Later, Majumder et al. (2019) demonstrated the use of GPT-based models for
personalized recipe generation, considering dietary needs, available ingredients, and
cultural preferences to generate customized cooking instructions. These approaches
play a key role in reducing food waste by encouraging the use of perishable ingredi-
ents before they expire.

For nutrition and sustainability tracking, open databases such as USDA FoodData
Central and Open Food Facts (2012) provide structured nutritional and environ-
mental data. Poore and Nemecek (2018) emphasized the importance of linking
ingredient-level data with life cycle assessments (LCA) to estimate a meals environ-
mental footprint, including carbon emissions and water usage.

Department of CSE 2 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

CHAPTER 3

METHODOLOGY AND PROPOSED SYSTEM

3.1 CNN (Convolutional Neural Network)

Convolutional Neural Networks (CNNs) are a specialized class of deep learning ar-
chitectures designed specifically for visual data such as images and video frames.
CNNs eliminate the need for handcrafted feature extraction by learning patterns
directly from raw pixel values. During training, the network automatically devel-
ops hierarchical features: the early layers detect low-level structures such as lines,
corners, and edges, while deeper layers learn complex and abstract patterns such
as facial landmarks, emotional expressions, or object categories. This hierarchical
learning ability makes CNNs highly effective in tasks such as facial recognition, med-
ical imaging analysis, autonomous driving, and emotion recognition, where precision
and pattern understanding are essential.

When an image is given as input, the CNN first processes it through convolutional
layers, where small filters slide over the image to detect low-level patterns such as
edges, colors, and textures. These feature maps are then passed through activation
functions like ReLU to introduce non-linearity, followed by pooling layers that re-
duce the spatial size and retain only the most important features, making the model
more efficient and less sensitive to variations in lighting, angle, or background. As
the network goes deeper, the extracted features become more abstract, enabling
the CNN to recognize complex food characteristics such as shape, ingredient ar-
rangement, and dish structure. Finally, the flattened features are passed into fully
connected layers, where the network combines all learned information and classifies
the image into its corresponding food category. This entire process allows the CNN
in the proposed system to achieve high accuracy in Indian food recognition, forming
the foundation for recipe generation, nutritional analysis, and sustainable cooking
recommendations.

3.1.1 Input Layer

The input layer is the entry point of the CNN, where the image data is initially fed
into the network. In facial expression recognition, the input is usually a grayscale

Department of CSE 3 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

image of a fixed dimension such as 48 × 48 pixels to ensure consistency across all

samples. This layer itself does not perform any computation or transformation;
rather, it simply forwards the raw pixel values to the next processing layer. Stan-
dardizing the input size ensures that the network receives uniform data regardless
of variations in the original image resolution. Additionally, preprocessing steps such
as normalization or pixel scaling are often applied before or at this stage to enhance
convergence and improve model stability.

3.1.2 Convolutional Layer

The convolutional layer is the foundational and most critical component of a Convo-
lutional Neural Network, as it is responsible for extracting meaningful visual features
from the input image. Instead of processing the image as a whole, this layer applies
small learnable matrices called filters or kernels, which slide across the image in a
process known as convolution. At each step of this movement (called a stride), the
filter performs an element-wise multiplication with the pixels beneath it and pro-
duces a single numerical value. This process generates what is known as a feature
map, which highlights the presence and location of specific visual patterns in the
image.

Each filter is trained to detect a specific feature — for example, vertical or hori-
zontal edges, corners, gradients, curves, textures, or fine details such as the outline
of eyes or the curvature of lips. The use of multiple filters in a single convolutional
layer allows the network to learn a rich set of visual characteristics. Early layers
typically detect low-level features such as sharp edges, lines, and simple shapes,
while deeper layers progressively extract high-level and abstract features, including
facial landmarks, emotional cues, and expression-specific patterns. This hierarchi-
cal learning ability enables CNNs to analyze complex structures without requiring
manually engineered feature descriptors.

A feature map is generated in the convolutional layer when a filter or kernel slides
over the input image and performs a mathematical operation known as convolution.
The input image is represented as a matrix of pixel values, and the filter, which is
usually much smaller, moves across the image region by region. At each position,
the values in the filter are multiplied element-wise with the corresponding pixel val-
ues beneath it, and the resulting values are summed to produce a single number.
This number represents how strongly the filter detects a particular pattern—such as
a vertical edge, corner, or texture—at that location. As the filter continues sliding

Department of CSE 4 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

across the image, repeating this process, a new matrix is formed, which is referred
to as the feature map. The activation function, typically ReLU, is then applied to
remove negative values and retain only meaningful activations. High values in the
feature map indicate that the visual feature targeted by the filter is present in that
region of the image, while low or zero values mean the feature is absent. Through
this process, the CNN learns to detect important visual patterns automatically, and
as more layers are added, feature maps evolve from capturing simple low-level edges
to complex high-level representations such as facial components or emotional cues.

Figure 3.1: CNN-Convolutional Layer.

3.1.3 Activation Function (ReLU)

After the convolution operation generates feature maps, an activation function is

applied to introduce non-linearity into the network and enable it to learn complex
patterns. One of the most widely used activation functions in CNNs is the Recti-
fied Linear Unit (ReLU). The ReLU function applies a simple transformation: it
sets all negative values in the feature map to zero while retaining positive values
unchanged. Although mathematically simple, this function plays a crucial role in
deep learning. By removing negative activations, ReLU ensures that only meaning-
ful features—those that strongly activate the filter—are carried forward to the next
layer.

More importantly, ReLU enables the network to model non-linear relationships,

which is essential for tasks such as facial expression recognition where subtle differ-

Department of CSE 5 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

ences in muscle movement define different emotions. Without an activation function

like ReLU, a CNN would behave like a purely linear system, regardless of the number
of layers, and would fail to learn complex hierarchical representations. Additionally,
ReLU helps mitigate optimization challenges such as the vanishing gradient prob-
lem, allowing gradients to flow more effectively during backpropagation. This leads
to faster training, deeper model architectures, and improved performance in recog-
nizing fine-grained emotional cues, making ReLU integral to modern CNN-based
recognition systems.

Figure 3.2: CNN-ReLU Activation.

3.1.4 Pooling Layer

The pooling layer plays an essential role in CNN architecture by reducing the spa-
tial dimensions of the feature maps generated by the convolutional layers while
preserving the most meaningful information extracted from the input image. By
downsampling the feature maps, pooling not only decreases the computational load
but also reduces the number of trainable parameters, making the model more effi-
cient and faster to train. The most commonly used technique is Max Pooling, where
the filter scans small regions and selects the single highest pixel value within each
region. This ensures that the strongest or most relevant feature — for example, a
sharp edge, a facial wrinkle, or a contour — is retained while less important data is
discarded.

Pooling also introduces an important characteristic known as translation invari-

ance, meaning that the model becomes less sensitive to minor shifts, rotations, or
distortions in the input image. For instance, if a person slightly moves their head or
changes facial angle, the extracted features remain recognizable even though their

Department of CSE 6 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

exact pixel position has changed. This robustness is crucial in real-world applica-
tions like facial expression recognition, where facial features may not always align
perfectly with the training data. Additionally, pooling helps prevent overfitting
by reducing the level of detail the network memorizes, encouraging it instead to
generalize patterns. As a result, even when emotion samples vary due to light-
ing, background, camera distance, or slight facial movement, the network can still
identify expressions accurately and consistently.

Figure 3.3: CNN-Pooling Layer.

3.1.5 Fully Connected Layers

After the convolutional and pooling layers extract increasingly complex visual fea-
tures from the input food images, the fully connected layer appears toward the final
stage of the CNN architecture. At this point, the multidimensional feature maps are
flattened into a one-dimensional vector so they can be processed like a traditional
neural network. Each neuron in the fully connected layer is connected to every
neuron in the previous layer, allowing the network to combine all the spatial and
semantic features learned throughout the convolutional layers. This dense represen-
tation helps the model perform high-level reasoning and make accurate classification
decisions. Finally, the output layer uses the Softmax activation function to convert
the computed scores into class probabilities. Based on these probabilities, the model
identifies the input image as one of the predefined Indian food categories, enabling
the system to provide accurate recipe suggestions, nutritional analysis, and sustain-
ability recommendations.

Department of CSE 7 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

Figure 3.4: CNN-Fully Connected Layer .

3.2 NLP (Natural Language Processing)

Natural Language Processing (NLP) forms a core component of the proposed system
by enabling intelligent understanding, interpretation, and generation of human-like
cooking instructions. Once the food item or ingredients are identified, the NLP
module processes user inputs—such as dietary restrictions, preferred cuisine type,
cooking time limitations, and ingredients available in the pantry—to generate mean-
ingful and coherent recipes. The GPT-based language model used in this system
employs advanced NLP techniques, including tokenization, contextual embedding,
semantic parsing, and sequence prediction. These techniques allow the system to
break down user prompts into structured elements, understand relationships be-
tween ingredients and cooking actions, and generate step-by-step instructions that
are logically consistent and easy to follow. The model leverages large-scale pre-
training, enabling it to capture linguistic nuances, culinary terms, and ingredient
interactions without requiring manual rule creation. Additionally, NLP allows the
system to analyze ingredients that are about to expire and suggest recipes that use
them efficiently, reducing household food waste. The module also offers ingredient
substitutions by understanding semantic similarity between food items, ensuring
flexibility in meal preparation. By combining language understanding with contex-
tual reasoning, the NLP component transforms simple user inputs into personalized,
sustainable, and nutritionally meaningful cooking recommendations.

Department of CSE 8 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

3.3 GPT-based Language Model (Generative Recipe Gen-

eration)
The proposed system uses a GPT-based language model to intelligently generate
personalized and context-aware recipes based on the food items detected by the
CNN or RoboFlow models. GPT (Generative Pre-trained Transformer) is a state-
of-the-art Natural Language Processing architecture that excels at understanding
user inputs, predicting sequences, and generating human-like text. In this system,
the GPT model analyzes user-provided information such as available ingredients,
dietary restrictions, cuisine preferences, cooking time, and nutritional goals. Using
techniques such as tokenization, contextual embedding, and attention mechanisms,
the model identifies relationships between ingredients and cooking actions, allow-
ing it to generate coherent, step-by-step recipe instructions. It can also suggest
ingredient substitutions based on semantic similarity, helping users make flexible
choices while minimizing food waste. Unlike traditional rule-based systems, the
GPT model does not rely on predefined templates; instead, it uses deep contextual
understanding to produce creative, natural, and diverse recipe ideas. This capability
makes the recipe generation module highly adaptable, user-oriented, and effective
for promoting sustainable cooking practices.

3.4 RoboFlow Pre-trained Image Detection API

The RoboFlow Pre-trained Image Detection API is integrated into the proposed
system to extend food recognition capabilities beyond Indian cuisine and support
the identification of a wide range of international dishes and ingredients. RoboFlow
provides powerful computer vision models that are already trained on large, diverse,
and high-quality food datasets, eliminating the need for manual data collection,
annotation, and additional training. When a user uploads an image, the API pro-
cesses it through its pre-trained object detection models, extracts relevant visual
features, and returns accurate predictions along with confidence scores. This allows
the system to recognize global food items in real time, even if they are not part of
the locally trained CNN model. RoboFlow also handles variations in lighting, back-
ground, angles, and food presentation, ensuring robust performance in real-world
conditions. By relying on this API, the platform achieves scalability, reduces com-
putational workload, and ensures quick integration of new food categories without
frequent model updates. The inclusion of RoboFlow enhances the overall versatility
and global applicability of the sustainable cooking system.

Department of CSE 9 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

3.5 Softmax Classification

The Softmax function is an activation function commonly used in the output layer
of neural networks for multi-class classification problems. It transforms the raw
output values (also called logits) produced by the network into probabilities that
sum up to one. This allows the model to represent how likely each class is to be
the correct answer. The Softmax function works by taking the exponential of each
output value and then dividing each exponential by the sum of all exponentials
across the classes. As a result, larger input values receive higher probabilities, while
smaller ones get lower probabilities. For example, if the model produces scores for
three possible emotions, Softmax converts them into percentages showing how con-
fident the model is about each emotion. It is widely used in tasks such as facial
emotion recognition, image classification, and text categorization, where the model
must select one label from several possible options. Softmax also pairs naturally
with categorical cross-entropy loss, making it an essential component in training
deep learning models that predict distinct categories.

e zi
Softmax(zi ) = P zj (3.1)
je

where

• zi is the score (logit) for class i,

• e is the exponential constant (approximately 2.71828),

• the denominator j ezj normalizes the outputs so that the resulting values
P

form a probability distribution (i.e. they sum to 1).

3.6 PROPOSED SYSTEM

3.6.1 Data Collection

For this project, datasets were gathered to address two primary objectives: In-
dian food classification using a CNN model and international food detection using
external APIs.

For Indian food recognition, a curated dataset containing high-resolution images of

common Indian dishes was collected. The dataset was sourced from publicly avail-
able repositories such as Kaggle and other open-access image datasets. Care was
taken to ensure that the dataset included a wide variety of dishes representing dif-
ferent regions of India, such as idli, dosa, biryani, samosa, and paneer-based recipes.

Department of CSE 10 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

Images were carefully filtered to include diverse backgrounds, lighting conditions,

and angles to improve model generalization. Each image was manually annotated
with its corresponding food label to create a structured and reliable training set.
This dataset was then split into three subsets: 70 percent for training, 20 percent for
validation, and 10 percent for testing. This division allowed the Convolutional Neu-
ral Network (CNN) model to learn effectively while minimizing overfitting. After
preprocessing, such as resizing and normalization, the dataset was ready for use in
model training. The final model achieved 92 percent accuracy in identifying Indian
food categories.

For international food detection, instead of manually collecting and labeling a mas-
sive dataset, the RoboFlow API was utilized. RoboFlow provides pre-trained models
and datasets capable of detecting a wide range of global cuisines and ingredients.
By integrating RoboFlow into the system, the platform could instantly recognize in-
ternational dishes without the need for additional data labeling or training, making
it highly scalable and efficient.

The combination of a locally trained Indian food dataset and the RoboFlow API
ensured comprehensive coverage of both regional and international cuisines. This
hybrid approach allowed the system to provide accurate food recognition results,
forming the foundation for subsequent features such as recipe generation, nutri-
tional analysis, and waste tracking. By emphasizing diverse and high-quality data
sources, the platform effectively supports sustainable cooking practices through re-
liable AI-driven insights.

3.6.2 Data Preprocessing

Preprocessing is a crucial step in the development of any AI or machine learn-

ing model, as it ensures that the input data is clean, consistent, and suitable for
analysis. In this project, preprocessing was applied to both image data and textual
data, forming the foundation for accurate food recognition, recipe generation, and
nutritional tracking. The main goal of preprocessing was to optimize the dataset for
the Convolutional Neural Network (CNN) used for Indian food classification and to
prepare real-time inputs for smooth functioning of the integrated modules.

For image data, the preprocessing began with cleaning and organizing the collected
Indian food dataset. Duplicate, low-quality, and blurred images were removed to
maintain data quality. Each image was resized to a fixed resolution of 224 x 224

Department of CSE 11 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

pixels, ensuring uniformity and compatibility with the CNN models input layer.
Pixel normalization was then performed by scaling pixel values between 0 and 1,
which helps in faster convergence during model training and reduces computational
complexity.

To improve the models generalization capability and handle real-world variations,

data augmentation techniques were applied. These included random rotations, hor-
izontal and vertical flipping, zooming, cropping, and brightness adjustments. Aug-
mentation not only artificially increased the size of the dataset but also helped the
CNN adapt to diverse conditions such as different camera angles, lighting variations,
and plating styles. This step was essential for achieving the final accuracy of 92 per-
cent during testing.

The dataset was then split into three subsets: 70 percent for training, 20 percent for
validation, and 10 percent for testing. This structured split ensured that the model
could be trained effectively, tuned using validation data, and evaluated on unseen
test images to measure real-world performance.

For text data, preprocessing was applied in the recipe generation module. Nat-
ural Language Processing (NLP) techniques such as tokenization, stemming, and
dependency parsing were used to break down ingredient names, quantities, and
cooking instructions into structured tokens. This allowed the GPT-based model to
understand user preferences, identify relationships between ingredients, and gener-
ate coherent, step-by-step recipes.

Overall, preprocessing ensured that both image and text inputs were accurate, con-
sistent, and ready for model training and deployment. This step significantly en-
hanced the reliability and performance of the platform, forming a solid foundation
for sustainable cooking applications.

3.6.3 Feature Extraction

Feature extraction is a crucial stage in the development of the AI-powered sus-

tainable cooking platform, as it enables the system to identify key characteristics
of food images and textual data for accurate classification, recipe generation, and
nutritional analysis. By extracting meaningful features, the platform can process
complex data inputs and transform them into structured information that drives
reliable decision-making.

Department of CSE 12 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

For food recognition, feature extraction is performed using a Convolutional Neu-

ral Network (CNN) trained on the Indian food dataset. CNNs are highly effective
for image-based tasks because they automatically detect and learn visual patterns
such as shapes, textures, and colors. The process begins with convolutional lay-
ers, which apply filters to the input image to capture low-level features like edges,
contours, and color gradients. As the network goes deeper, higher-level features
such as ingredient patterns, dish shapes, and plating styles are extracted. Pooling
layers are then used to reduce the spatial dimensions of the data while preserving
the most important features, improving computational efficiency and reducing over-
fitting. Finally, fully connected layers combine these extracted features to produce
a classification output, identifying the dish with high accuracy. During training,
the CNN achieved 92 percent accuracy on Indian food images, demonstrating its
effectiveness in capturing distinctive features of regional cuisines.

For international food recognition, the RoboFlow API provides pre-trained feature
extraction models. These models analyze uploaded images and extract relevant fea-
tures without requiring additional local training, making the system scalable for a
wide range of global cuisines and ingredients.

In the recipe generation module, feature extraction is applied to textual data using
Natural Language Processing (NLP) techniques. The GPT-based model performs
tokenization, breaking down recipe text and ingredient descriptions into smaller
units for analysis. Dependency parsing is then used to understand the relationships
between ingredients, quantities, and cooking actions.

The combination of visual and textual feature extraction creates a powerful multi-
modal [Link] features ensure precise food recognition,while textual fea-
tures enable intelligent recipe customization and nutritional [Link] synergy
allows the platform to deliver accurate,personalized, and sustainability-focused rec-
ommendations to users,forming the foundation for smart,waste-reducing cooking
practices.

3.6.4 Data Labelling

Data labeling is a critical step in developing accurate and reliable machine learn-
ing models, especially for supervised learning tasks such as food image recognition.
The quality and consistency of labels directly affect the models ability to learn and
generalize to real-world scenarios. In this project, data labeling was carried out

Department of CSE 13 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

carefully for the Indian food dataset, which was used to train the Convolutional
Neural Network (CNN) model, while the RoboFlow API was leveraged for interna-
tional food recognition, where pre-labeled datasets were already integrated.

For Indian food classification, a curated dataset consisting of high-resolution im-

ages of popular dishes from diverse regions of India was collected. Each image was
manually inspected and assigned a single, clear label corresponding to its food cat-
egory, such as Idli, Dosa, Paneer Butter Masala, Biryani, or Samosa. The labeling
process was conducted using annotation tools like LabelImg and RoboFlow Annota-
tor, which allow images to be tagged and organized into structured datasets. Special
care was taken to maintain consistency, ensuring that similar dishes with slight vari-
ations were labeled uniformly to prevent confusion during model training.

The dataset was structured into directories for each food class, following a super-
vised learning approach where each folder represented a unique label. This format
enabled the CNN model to map image features to their respective dish categories
effectively. The labeled dataset was then divided into training (70 percent), valida-
tion (20 percent), and testing (10 percent) subsets. The training set was used to
teach the model, the validation set helped fine-tune hyperparameters, and the test
set evaluated the final models performance.

For international food detection, manual labeling was not required, as the RoboFlow
API provided access to pre-trained models that use globally recognized food datasets
with high-quality annotations. This integration significantly reduced the effort re-
quired for labeling and enabled the platform to instantly recognize a wide variety of
international cuisines and ingredients.

By combining manually labeled local data with pre-labeled global datasets, the
platform ensured a robust and scalable labeling framework. This hybrid labeling
approach laid a strong foundation for achieving 92 percent accuracy in Indian food
recognition while maintaining flexibility for international dish detection.

3.6.5 Training and Testing

Training and testing are vital stages in developing an accurate and reliable AI
model for food recognition and recipe generation. These steps ensure that the model
not only learns from the provided data but also performs well on unseen data, making
it robust and suitable for real world applications. In this project, the Convolutional

Department of CSE 14 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

Neural Network (CNN) was trained to classify Indian food items, while RoboFlow
pre trained models were used for detecting international cuisines.

The training phase began with the preprocessed and labeled Indian food dataset.
Images were divided into three subsets: 70 percent for training, 20 percent for val-
idation, and 10 percent for testing. During training, the CNN learned to identify
and extract hierarchical visual features from food images. The convolutional layers
captured low-level features such as edges, shapes, and textures, while deeper layers
learned complex patterns specific to different dishes. To enhance the models gener-
alization capabilities and prevent overfitting, techniques such as data augmentation,
dropout regularization, and batch normalization were employed. The training pro-
cess was conducted using the TensorFlow framework, with optimization handled by
the Adam optimizer and categorical cross-entropy loss function. Hyperparameters
such as learning rate, batch size, and number of epochs were fine-tuned using the
validation set to achieve optimal performance.

The testing phase evaluated the CNNs performance on completely unseen images
from the test set. Key metrics such as accuracy, precision, recall, and F1-score were
calculated to assess the models effectiveness. The trained CNN achieved 92 percent
accuracy in classifying Indian food items, indicating its strong ability to distinguish
between different dishes. The results also highlighted the models robustness in han-
dling variations in plating styles, lighting conditions, and backgrounds.

For international food recognition, additional training was not required as the RoboFlow
API provided pre-trained models capable of detecting a wide range of global cuisines.
These models were directly integrated into the system and tested using sample im-
ages to verify their reliability and compatibility with the platform.

By combining a locally trained CNN with pre-trained global models, the system
achieved comprehensive food recognition coverage. This solid foundation enabled
downstream modules, such as recipe generation and nutritional analysis, to function
accurately and effectively, supporting the platforms goal of promoting sustainable
cooking practices.

3.6.6 Personalised Recipe Generation

After the completetion of the food classification, the system uses a GPT model
based Natural Language Processing to generate custom recipes. The model runs

Department of CSE 15 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

through all the user input, such as recognised ingredients, dietary needs (vegan,
gluten free, low carb etc), culinary cultural preferences and products in the pantry
to generate individualised meal options. The GPT model predicts the next word
in a sentence and generates coherent cooking recipes. The model can also suggest
substitute ingredients by user choice, so you can be flexible and responsive during
cooking. All the recipes include step by step cooking instructions, cooking time,
serving size and substitute ingredients. With AI recipe generation the system en-
ables users to be more creative in the kitchen and reduces food waste by eating what
you have. The system is fine tuning its AI models with user input and real data to
make them accurate and relevant.

The food classification model gets more training data to classify new foods better.
The GPT driven NLP model used in recipe generation is updated incrementally by
user engagement, user preference and new food trends as shown in Figure 1. The
features for nutritional analysis and waste monitoring are updated by API to keep
the system up to date with latest nutritional guidelines and sustainability prac-
tices. All these incremental updates makes the AI chef more efficient, accurate and
relevant.

Figure 3.5: Functioning of natural language processing

Department of CSE 16 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

3.6.7 System Framework

The framework of the system is implemented in such a way that it is inte-

grated with various AI-powered elements that ensure sustainable cooking method-
ologies. The design starts from Ingredient Detection, which employs convolutional
neural networks (CNNs) for food images classification to provide proper identifica-
tion of ingredients. International food classification employs the RoboFlow API so
that classification functions can be increased for a vast array of cuisines worldwide.
Upon the determination of ingredients, the system utilizes a GPTbased NLP model
to create customized recipes based on individual tastes, dietary needs, and pantry
contents.

A Waste Tracking module is also used to enable users to record unused or ex-
piring ingredients to reduce food wastage by giving users customized meal ideas
that utilize these ingredients. The Nutrition Analysis is facilitated by the API
Ninjas platform, providing instantaneous information on meals’ macronutrient and
micronutrient content. In a systematic process of combining the elements, the sys-
tem provides a holistic and astute method to encourage sustainability as well as
healthy cooking as shown in Figure below.

Figure 3.6: Architectural system

Department of CSE 17 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

3.6.8 System Deployment and User Interface Design

The deployment of the AI-powered sustainable cooking platform was carefully

planned to ensure scalability, accessibility, and a smooth user experience. Once the
core modules including food recognition, recipe generation, nutritional tracking, and
waste monitoring were developed and tested, they were integrated into a single web-
based application using Streamlit. Streamlit was chosen for its simplicity, flexibility,
and ability to support interactive AI-powered applications with minimal overhead.

The deployment process began by packaging the trained Convolutional Neural Net-
work (CNN) model for Indian food recognition and integrating it with the RoboFlow
API, which handles international food detection. These modules were connected
to the Natural Language Processing (NLP) component, powered by a GPT-based
model, for generating personalized recipes and step-by-step cooking instructions.
External APIs, such as USDA FoodData Central and other nutrition and sustain-
ability databases, were integrated to provide real-time insights into nutritional values
and environmental impact. The entire system was hosted on a cloud-based platform
to ensure that it could handle multiple simultaneous user requests efficiently.

The user interface (UI) was designed with a focus on simplicity and usability, making
sustainable cooking accessible to a wide range of users. Built using Streamlit, the
interface allows users to interact seamlessly with the platform through an intuitive
dashboard. The UI enables users to upload images of ingredients or dishes directly
from their device camera or gallery. Once the image is processed, the system displays
the recognized food items along with nutritional information, carbon footprint, and
other sustainability metrics. Users can then generate customized recipes tailored to
their preferences and available ingredients.

Additional features, such as meal planning and waste tracking, are presented in
an organized layout to encourage consistent sustainable practices. Visual elements
such as progress bars, charts, and eco-impact indicators were included to make the
data more engaging and easy to interpret.

By combining robust backend processing with a clean, interactive front end, the sys-
tem provides a comprehensive solution for promoting sustainable cooking. The de-
ployment approach ensures that users can access advanced AI functionalities in real-
time, while the intuitive interface fosters regular engagement and behavior change
toward eco friendly and health conscious food habits.

Department of CSE 18 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

CHAPTER 4

RESULTS AND DISCUSSION

The proposed AI-powered sustainable cooking platform was evaluated compre-

hensively to measure its effectiveness in achieving its core objectives: accurate food
recognition, intelligent recipe generation, nutritional and sustainability tracking,
and overall usability for promoting eco-friendly cooking practices. The evaluation
focused on both the technical performance of the models and the user experience,
ensuring that the system provides meaningful real-world impact.

The first module evaluated was the Convolutional Neural Network (CNN) de-
veloped for Indian food classification. The CNN was trained on a curated dataset
consisting of high-resolution images of diverse Indian dishes. During training, the
model achieved rapid convergence due to effective preprocessing steps such as im-
age normalization and augmentation. After fine-tuning hyperparameters including
batch size, learning rate, and number of epochs, the final trained model achieved a
testing accuracy of 92 percent.

The CNN demonstrated high precision in identifying commonly consumed dishes

such as Idli, Dosa, Paneer Butter Masala, and Biryani. Data augmentation proved
crucial in improving the robustness of the model, enabling it to correctly classify
images under varying real-world conditions such as changes in lighting, plating style,
and background noise.

The performance was further analyzed using a confusion matrix, which highlighted
strong classification accuracy across most categories. Minimal misclassifications
were observed, primarily between visually similar dishes, such as Vegetable Pulao
and Biryani, or Chapati and Paratha. These errors suggest the need for a larger
and more diverse dataset to further improve model generalization. Nonetheless, the
overall results demonstrate the CNNs effectiveness as a reliable tool for automated
food recognition.

For international food detection, the platform integrated the RoboFlow API, which
provides access to pre-trained models trained on a vast global dataset. This elimi-

Department of CSE 19 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

nated the need for additional manual data labeling or training while ensuring that
the system could detect a wide variety of international cuisines and ingredients.
Testing with sample images confirmed that RoboFlow seamlessly integrated with
the CNN module, providing accurate and real-time results for dishes outside the
Indian dataset. This hybrid approach significantly expanded the systems coverage
and scalability.

The second major component evaluated was the recipe generation module, pow-
ered by a GPT-based Natural Language Processing (NLP) model. The module was
designed to generate personalized, context-aware recipes based on the identified in-
gredients and user preferences.

Qualitative testing was carried out by providing the system with different sets of
ingredients and dietary constraints. The generated recipes were coherent, easy to
follow, and culturally relevant. For example, when the identified ingredients in-
cluded rice, lentils, and vegetables, the system suggested sustainable dishes such as
Vegetable Khichdi or Stir-Fried Rice while also offering alternative suggestions for
missing or near-expiry items.

One of the most impactful features was the systems ability to provide ingredient
substitutions. This functionality directly supports the goal of reducing food waste
by ensuring that users can utilize available pantry items instead of discarding them.
By integrating external nutrition and sustainability databases, the system also dis-
played real-time nutritional values, such as calorie count and macronutrient com-
position, along with environmental impact metrics like carbon footprint and water
usage. This enabled users to make informed decisions, not only for their health but
also for environmental conservation.

The user interface (UI) was developed using Streamlit, chosen for its lightweight
design and interactivity. The dashboard was tested with a group of users to evalu-
ate ease of use and responsiveness. Users could upload images, view real-time food
recognition results, and navigate seamlessly between modules such as recipe gener-
ation, nutritional analysis, meal planning, and waste tracking.

Feedback collected during usability testing was overwhelmingly positive. Users ap-
preciated the platforms clean layout and visual elements such as progress bars, eco-
footprint indicators, and suggested meal plans. Trial users reported that the system

Department of CSE 20 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

encouraged them to rethink their cooking habits and provided practical strategies
for reducing food waste at home.

The results of this project highlight the effectiveness of combining AI, NLP,
and external APIs to create a practical solution for sustainable cooking. The CNN
achieved a strong accuracy of 92 percent, confirming its reliability in identifying
Indian dishes, while the RoboFlow API extended the systems capabilities to inter-
national cuisines without additional training overhead.

The GPT-based recipe generation module successfully delivered customized, waste-

conscious recipes, making the cooking experience both engaging and environmentally
responsible. Users were able to understand the nutritional value and environmental
footprint of their meals, encouraging more sustainable decision-making.

However, certain challenges were identified during testing. Misclassification between

visually similar dishes indicates the need for expanding and diversifying the dataset.
Additionally, the systems reliance on external APIs poses a limitation in environ-
ments with poor internet connectivity. Future improvements could include offline
functionality and IoT integration with smart kitchen devices for real-time inventory
management.

Despite these challenges, the platform demonstrates significant potential for real-
world adoption. It serves as a one-stop solution for food recognition, recipe plan-
ning, and sustainability tracking, addressing both household food waste and broader
environmental concerns. These findings confirm that AI can play a transformative
role in building sustainable food ecosystems, making eco-friendly cooking accessible
and practical for users worldwide.

Department of CSE 21 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

CHAPTER 5

ADVANTAGES AND DISADVANTAGES

5.1 ADVANTAGES
The proposed AI-powered sustainable cooking platform offers several advantages
that make it a powerful and practical solution for addressing food waste, promoting
healthy eating, and encouraging environmentally sustainable practices. By integrat-
ing deep learning, natural language processing (NLP), and real-time sustainability
tracking, system goes beyond traditional cooking and recipe applications to deliver
a comprehensive, intelligent solution.

One of the major advantages of this platform is its ability to reduce food waste
through intelligent ingredient management. The system identifies available pantry
items, including those nearing expiration, and generates recipes to ensure they are
used efficiently rather than discarded. It also provides creative suggestions for re-
purposing leftovers and offers ingredient substitutions, directly tackling one of the
most common causes of household food wastage.

Another benefit is personalized recipe generation, powered by a GPT-based NLP

module. Unlike generic cooking apps, this platform tailors its recipes to individual
dietary preferences, health goals, and cultural habits. The generated recipes include
step-by-step cooking instructions, sustainable substitutions, and time-saving tips,
making cooking simpler, healthier, and more eco-friendly.

The food recognition module, built using a Convolutional Neural Network (CNN),
achieved 92 percent accuracy in identifying Indian dishes. Integration with the
RoboFlow API further enhances this capability by detecting international cuisines
without the need for additional manual training, making the platform scalable for a
global [Link] a Streamlit-based interface, the platform provides an intuitive
and accessible user experience. By combining food recognition, recipe generation,
and real-time nutritional and environmental insights, it empowers users to make
informed, sustainable choices in their everyday cooking practices.

Department of CSE 22 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

5.2 DISADVANTAGES
While the proposed AI-powered sustainable cooking platform offers several inno-
vative features and benefits, it also has certain limitations and challenges that must
be addressed for its widespread adoption and long-term effectiveness.

One of the primary disadvantages is its dependence on high-quality datasets. The

Convolutional Neural Network (CNN) used for Indian food recognition achieved 92
percent accuracy, but its performance may decline when dealing with low-quality,
blurred, or uncommon dish images. Misclassifications can occur, especially between
visually similar foods, such as Vegetable Pulao and Biryani, which could affect recipe
generation and ingredient management. Expanding and diversifying the dataset is
essential to overcome this limitation.

Another drawback is the platforms reliance on external APIs, such as RoboFlow

for international food detection and nutritional databases for real-time insights. In
regions with poor internet connectivity, this reliance can limit functionality, making
some features unavailable or slow to respond. Additionally, frequent API updates
or access restrictions could disrupt system performance over time.

Privacy concerns also pose a challenge. Since the platform collects user data, in-
cluding dietary preferences and uploaded images, there is a risk of data breaches or
misuse if proper security measures are not implemented.

The system may also face adoption barriers among users unfamiliar with digital
platforms or hesitant to change their cooking habits. Finally, integration with real-
time sustainability tracking requires accurate environmental data, which may not
always be region-specific, potentially reducing the reliability of eco-footprint metrics.

Despite these challenges, these disadvantages can be addressed through continuous

improvements, stronger datasets, and enhanced system design.

Department of CSE 23 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

CHAPTER 6

FUTURE SCOPE

The current AI-powered sustainable cooking platform has achieved strong results
in food recognition, recipe generation, and waste tracking, but there are many op-
portunities to enhance its functionality and broaden its impact. The Convolutional
Neural Network (CNN) developed for Indian food classification achieved a notable
92 percent accuracy, while the RoboFlow API successfully handled international
food detection. However, in the future, this module can be expanded by training
on larger and more diverse datasets, covering a wider range of cuisines and regional
variations. This will improve the systems ability to identify visually similar dishes
and reduce dependency on external APIs, making the model more robust and inde-
pendent.

The recipe generation module, powered by a GPT-based Natural Language Pro-

cessing (NLP) model, can also be scaled further. Future versions can leverage larger
language models (LLMs) to generate more context-aware and nutritionally optimized
recipes. Features such as voice-based interaction and hands-free cooking assistance
can be added to improve accessibility for users in real-time cooking scenarios.

The Streamlit interface, while user-friendly, can be developed into a mobile ap-
plication with offline capabilities, ensuring seamless access even in areas with poor
internet connectivity. Additionally, integrating the system with IoT-enabled smart
kitchen devices such as smart refrigerators and weight sensors would allow automatic
pantry tracking, expiration date monitoring, and optimized grocery planning.

Beyond individual households, the platform can be scaled for restaurants, food de-
livery services, and NGOs, helping reduce food waste at an industrial level. Adding
multilingual support will make it accessible to diverse user groups across different
regions and cultures.

By implementing these advancements, the platform can evolve into a global solution
for food sustainability, contributing to environmental protection, waste reduction,
and healthier lifestyles worldwide.

Department of CSE 24 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

CHAPTER 7

CONCLUSION

The seminar on Sustainable Cooking with AI: Generative Recipes and Smart
Nutrition Assistance emphasized how artificial intelligence can serve as a practical
solution to pressing issues such as food waste, poor meal planning, and unhealthy
dietary habits. The presented system successfully integrates food classification, per-
sonalized recipe generation, nutrition assistance, and waste tracking into a single
user-friendly platform. By employing convolutional neural networks (CNNs) for ac-
curate food recognition, GPT-based natural language models for recipe creation,
and external APIs for real-time nutritional insights, the platform bridges advanced
technology with everyday cooking practices.

The results are promising, with the food classification model achieving an accuracy
of 92 percent and demonstrating robustness across varied real-world conditions such
as lighting, presentation, and ingredient diversity. The addition of the RoboFlow
API further extends its usability to international cuisines, making it inclusive for
users across different cultural contexts. The recipe generation module not only
adapts to dietary preferences such as vegan, gluten-free, or low-carb diets but also
suggests ingredient substitutions, ensuring flexibility and minimal food wastage.
Waste tracking and meal planning features encourage mindful consumption, helping
households reduce unnecessary purchases and use ingredients efficiently.

User feedback validates the effectiveness of the system, highlighting ease of use,
design intuition, and satisfaction with the meal planning experience. Beyond tech-
nical performance, the platform has proven its ability to promote eco-friendly living
by aligning daily food practices with sustainability goals.

Looking ahead, there remains scope for improvement, such as refining ingredient
detection for leftover tracking, expanding global cuisine coverage, and integrating
calorie and diet management tools. These additions would make the system even
more comprehensive. Overall, this work demonstrates the transformative potential
of AI in making cooking not only smarter and healthier but also more sustainable,
supporting both individual well-being and environmental conservation.

Department of CSE 25 College of Engineering Trikaripur

Seminar Report 2025 Sustainable Cooking with AI

BIBLIOGRAPHY

[1] Sheng, G., Min, W., Zhu, X., Xu, L., Sun, Q., Yang, Y., Wang, L.,
and Jiang, S. (2024). A Lightweight Hybrid Model with Location-Preserving
ViT for Efficient Food Recognition. Nutrients, 16(2), 200.

[2] Kansaksiri, P., Panomkhet, P., and Tantisuwichwong, N. (2023). Smart

Cuisine: Generative recipe and ChatGPT powered nutrition assistance for sus-
tainable cooking. Procedia Computer Science, 225, 2028-2036.

[3] Asha, P. N., Ramachandra, H. V., Sowmya Somnath, C., Subra-

manya, S. G., and TY, S. (2024). NLPANTRY: AI-POWERED RECIPE
CREATION. ACTA SCIENTIAE, 7(1), 486-495.

[4] Jabeen, H. (2024). Semantics for Culinary Health Care. In Roles and Chal-
lenges of Semantic Intelligence in Healthcare Cognitive Computing (pp. 51-67).
IOS Press.

[5] Peng, Y. (2024). SuperCook: Design Interactive Culinary Experiences with AI

Technology (Master’s thesis, Rochester Institute of Technology).

[6] Hannon, B., Kumar, Y., Li, J. J., and Morreale, P. (2024). Chef Dalle:
Transforming Cooking with Multimodal AI.

[7] Razzaq, M. S., Maqbool, F., Ilyas, M., and Jabeen, H. (2023).
EvoRecipes: a generative approach for evolving context-aware recipes. IEEE
Access.

[8] Ruiz-Rincon, S., and Galpin, I. (2024). SnapChef : AI-powered Recipe

Suggestions.

[9] Al-Sarayreh, M., Reis, M. G., Carr, A., and dos Reis, M. M. (2023).
Inverse design and ai/deep generative networks in food design: a comprehensive
review. Trends in Food Science and Technology, 138, 215- 228.

Department of CSE 26 College of Engineering Trikaripur