To enable real-time interpretation of Indian Sign Language (ISL), our system adopts a Hybrid CNN-Transformer Detection Framework. Standard convolutional backbones excel at extracting localized fine-grained features—such as finger shapes, wrist orientations, and sharp joint edges—while global context blocks (Multi-Head Self-Attention) model cross-spatial relationships across the hand gesture area. By unifying a lightweight YOLO convolutional backbone with attention mechanics, the model retains low computational complexity (FLOPs) while effectively capturing multi-scale spatial features across 32 unique ISL gesture classes.
 |
 |
| Practical exmaple 1 | Practical example 2 |
To ensure robustness under real-world visual variations (such as diverse lighting, background clutter, and subtle hand rotation variations), the model was trained using domain-tailored data augmentations:
- Spatial Transformations: Moderate rotation (±15°), translation (0.1), scale (0.5), shear (2.0), and horizontal flipping (
fliplr=0.5) to accommodate left- and right-handed sign variations without breaking gesture symmetry. - Color & Noise Augmentations: HSV color jittering (
hsv_h=0.015,hsv_s=0.5,hsv_v=0.4) and minor random erasing (erasing=0.1) to simulate non-ideal illumination and partial occlusions without masking distinct finger configurations. - Optimization: The network was optimized using AdamW (
lr0=0.001,weight_decay=0.0005) over 50 epochs with early stopping (patience=15), achieving fast convergence and mitigating over-fitting risk.
| Metric | Value | Description |
|---|---|---|
| Mean Precision (P) | 97.97% (0.9797) | Measures the proportion of correctly predicted Indian Sign Language (ISL) gesture detections among all positive predictions. |
| Mean Recall (R) | 99.50% (0.9950) | Indicates the model's ability to successfully detect and localize all ground-truth ISL hand gestures. |
| mAP @ IoU = 0.50 | 99.28% (0.9928) | Mean Average Precision evaluated at an Intersection over Union (IoU) threshold of 0.50, representing overall detection performance. |
| mAP @ IoU = 0.50:0.95 | 71.99% (0.7199) | Stringent Mean Average Precision averaged across IoU thresholds from 0.50 to 0.95, reflecting localization accuracy under stricter evaluation criteria. |
| Notebook Link - https://www.kaggle.com/code/sayaksamanta/indian-sign-language-detction