PROBLEM STATEMENT

Communication is a basic human need; however, hearing-

impaired individuals often face difficulties in expressing
themselves to the general public due to the lack of understanding
of sign language. Most people are not trained in sign language,
which creates a communication gap and leads to social
exclusion.

How can we help hearing-impaired people and the general

public to communicate effectively using sign language so that
real-time translation of sign gestures into readable text can
reduce communication barriers and promote social
inclusion?

This project aims to address this problem by developing a real-

time sign language (static image) to text recognition system
using Artificial Intelligence and Computer Vision techniques.

1
 STAKEHOLDERS

The main users and stakeholders of this project are:

• Hearing-impaired people
• General public
• Educational institutions
• Social inclusion initiative

2
 OBJECTIVES

The objectives of this project are:

• To recognize American Sign Language (ASL)
alphabets in real time using a webcam
• To convert hand sign gestures into readable text
output
• To reduce communication barriers for hearing-
impaired individuals
• To apply computer vision and deep learning concepts
practically
• To build an efficient and accurate AI-based
recognition system

3
 AI DOMAIN USED

The project mainly uses the following AI domains:

• Computer Vision – for detecting and processing hand
gestures from images
• Machine Learning / Deep Learning – for training a
model to classify sign language alphabets
AI is required for this problem because manual recognition of
hand gestures is not feasible in real time. AI models can
automatically learn visual patterns from large datasets and
accurately recognize sign gestures.

4
 DATA USED
• Type of Data: Image data (static hand gesture images)
• Dataset Source: American Sign Language (ASL)
dataset from Mendeley
• Data Description:
 Alphabets from A to Z
 Each letter contains 1000 images
 750 left-hand images and 250 right-hand images
per letter
 Only 60 images per letter were used for training

Ethical Considerations
• The dataset is publicly available and ethically sourced
• No personal or private information is used
• Bias may occur due to background and lighting variations,
which is acknowledged

5
 AI PROJECT CYCLE
IMPLEMENTATION

The development of this project follows the standard AI Project

Cycle as suggested in the CBSE AI Projects Cookbook. Each
stage is explained in detail below.
a) Problem Scoping
Problem scoping is the first and most important stage of the AI
Project Cycle. In this stage, the real-world problem was
identified and clearly defined. The main issue observed was the
communication gap between hearing-impaired individuals and
the general public due to lack of sign language knowledge. This
gap often leads to misunderstanding, dependence on others, and
social exclusion.
The scope of the project was limited to static American Sign
Language (ASL) alphabet recognition to keep the system
simple, accurate, and suitable for a Class XII level project. The
expected outcome was a system that could recognize hand
gestures in real time and convert them into readable text.
b) Data Acquisition
In this stage, data required for training the AI model was
collected. The dataset used was the American Sign Language
(ASL) alphabet dataset from Mendeley. The dataset contains
images of hand gestures representing alphabets from A to Z.

6
Each alphabet contains 1000 images, with 750 right-hand
images and 250 left-hand images. To reduce computational load
and training time, only 60 images per alphabet were selected
for training and testing. The dataset was well-organized into
folders, with each folder representing one alphabet.
c) Data Exploration
Data exploration was carried out to understand the structure and
quality of the dataset. The images were analyzed for:
• Hand orientation (left and right hand)
• Background uniformity
• Lighting conditions
• Image size and clarity
It was observed that background and lighting variations could
affect prediction accuracy. Hence, preprocessing steps were
planned to minimize these effects. This stage helped in
identifying limitations such as background bias and the need for
better lighting control during real-time prediction.
d) Modelling
For modelling, a deep learning approach was used.
MobileNetV2, a lightweight and efficient Convolutional Neural
Network (CNN), was selected because it provides high accuracy
with low computational cost, making it suitable for realtime
applications.
The dataset was split into:

7
 • 80% training data • 20% testing data
The model was implemented using TensorFlow and Keras.
During training, the model learned important features such as
hand shape, finger position, and gesture orientation.
e) Evaluation
The trained model was evaluated using accuracy as the only
evaluation metric, which is appropriate for a multi-class
classification problem at the school level. The model achieved
an accuracy of 96–97% on the test dataset. Real-time evaluation
was performed using a webcam to verify practical performance

PROPOSED SYSTEM
8
The proposed system is a real-time static sign language to text
recognition system designed to assist hearing-impaired
individuals in communicating with the general public. The
system uses a webcam to capture hand gestures and applies
computer vision and deep learning techniques to recognize
American Sign Language (ASL) alphabets.
System Working
The working of the system can be explained in the following
steps:
Step 1: Input Acquisition
A webcam continuously captures real-time video frames of the
user performing ASL hand gestures. Each frame acts as an input
image to the system.
Step 2: Preprocessing
The captured frames are processed using OpenCV.
Preprocessing includes resizing the image to a fixed size
compatible with MobileNetV2, normalization of pixel values,
and removal of unnecessary noise. This step ensures uniformity
in input data and improves prediction accuracy.
Step 3: Feature Extraction and Classification
The preprocessed image is passed to the trained MobileNetV2
model. The model extracts important features such as finger
position, hand shape, and orientation. Based on these features,
the model classifies the gesture into one of the ASL alphabet
classes (A–Z).
9
Step 4: Output Generation
The predicted alphabet is displayed as readable text on the
screen. This enables instant understanding of the sign made by
the user.
Webcam → Frame Capture → Preprocessing → MobileNetV2
Model → Text Output

TOOLS / TECHNOLOGIES USED

10
• Python
• OpenCV
• NumPy
• TensorFlow
• Keras
• MobileNetV2 (pre-trained model)

RESULT

11
The model successfully predicts ASL alphabets in real
time using a webcam. The output text corresponding to
the detected hand sign is displayed on the screen. The
system performs well under controlled lighting and
background conditions.

TESTING and EVALUATION

12
Testing and evaluation were carried out to verify both the
accuracy and real-time usability of the system.
Testing Methodology
• The dataset was divided into 80% training data and 20%
testing data.
• The trained model was tested on unseen images to evaluate
its generalization ability.
• Real-time testing was performed using a webcam under
different lighting and background conditions.
Evaluation Metric
Only accuracy was used as the evaluation metric, as it is
sufficient and appropriate for a multi-class classification
problem at the school level.
Performance : The model achieved an accuracy of 96–97% on
the test [Link] system successfully recognized most ASL
alphabets in real time.

Limitations Observed
• Performance decreases in poor lighting conditions
• Background bias affects prediction accuracy
These observations indicate areas where further improvement is
possible.

SDG MAPPING
13
This project aligns with:
SDG 10 – Reduced Inequalities
By enabling communication for hearing-impaired
individuals, the system promotes equality and social
inclusion.

CONCLUSION
14
This project successfully demonstrates the practical
application of Artificial Intelligence and Computer Vision
in solving a real-world problem. The developed sign
language to text recognition system effectively recognizes
static ASL alphabets and converts them into readable text
in real time.
The use of the MobileNetV2 model helped achieve high
accuracy while maintaining efficiency suitable for real-
time usage. The project highlights how AI can play an
important role in promoting accessibility and social
inclusion for hearing-impaired individuals.
Overall, the project met its objectives and provided
valuable learning experience in AI model development,
dataset handling, and real-time computer vision
applications.

FUTURE SCOPE

15
Future improvements can include:
• Dynamic gesture recognition
• Speech output for recognized text
• Improved accuracy with larger datasets
• Background bias removal and lighting adjustment
• Mobile or web application integration

INDIVIDUAL REFLECTION

16
Working on this AI Capstone Project was a valuable
learning experience. I gained a deeper understanding of
computer vision concepts, machine learning models, and
real-time image processing techniques.
One of the major challenges faced was handling a large
dataset and managing long training times. Another
difficulty was achieving accurate real-time predictions
due to background and lighting variations. Overcoming
these challenges improved my problem-solving skills and
technical confidence.
This project enhanced my interest in Artificial Intelligence
and motivated me to explore advanced topics such as
dynamic gesture recognition and speech-based AI systems
in the future.

BIBLIOGRAPHY
 ASL Dataset – Mendeley
17
 TensorFlow Documentation
 Keras Documentation
 OpenCV Python Documentation
 CBSE AI Project Cookbook

18