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
| Property | Value |
|---|---|
| Architecture | Custom CNN |
| Framework | TensorFlow / Keras |
| Input | 128 × 128 RGB |
| Classes | 29 |
| Model format | Keras .keras |
Performance
| Metric | Result |
|---|---|
| Validation Accuracy | 99.78% |
| Test Accuracy | 99.7241% |
| Precision | 99.7282% |
| Recall | 99.7241% |
| F1 Score | 99.7243% |
| Top-3 Accuracy | 99.9885% |
| External Accuracy | 100% |
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
- Downloads last month
- 9