signavision-cnn

SignaVision Custom CNN

SignaVision is a custom Convolutional Neural Network developed for American Sign Language (ASL) alphabet recognition.

The model classifies RGB images into 29 classes:

• A-Z
• delete
• nothing
• space

Model Details

Performance

The internal test set contains 8,700 images.

The external evaluation contains 28 separate sample images.

Classes

A B C D E F G H I J K L M
N O P Q R S T U V W X Y Z
del nothing space

Architecture

The model uses four convolutional blocks:
Input: 128 × 128 × 3
        ↓
Conv2D 32
        ↓
Conv2D 64
        ↓
Conv2D 128
        ↓
Conv2D 256
        ↓
Global Average Pooling
        ↓
Dense 256
        ↓
Dropout
        ↓
29-Class Softmax

Batch normalization and ReLU activation are used throughout the convolutional feature extraction stages.

Intended Use

The model is intended for:

Educational demonstrations
ASL alphabet recognition
Computer vision experimentation
Real-time sign recognition prototypes
Artificial Neural Network coursework
Limitations

This model recognizes static hand-sign images.

It is not a complete continuous sign-language translation system.

Dynamic signs and sentence-level language understanding are outside the scope of this model.

Real-world performance can vary with:

Lighting
Camera quality
Background
Hand position
Distance from camera
Different users
Different hand orientations