Computer Vision
Sign Language Recognition
A small CNN that recognizes hand signs for 24 letters of the ASL alphabet from your webcam — ~50 ms per prediction. Plus a closer look at which signs the model confuses (N vs S, C vs O) and why.
Objectives
- 1Recognize letters from images with sign language symbols using CNNs.
- 2Identify pairs of conflicting symbols when making predictions.
- 3Compare performance of different CNN architectures with and without Batch Normalization.
Conclusions
- The CNN achieved over 99% accuracy on the test set for 24-class classification.
- The most confused sign pairs are N-S and C-O, despite not being visually similar.
- Data augmentation (rotation, zoom, shift) improved generalization by 3-5%.
- Batch normalization improved numerical stability but did not increase accuracy.
Technologies
- TensorFlow
- Keras
- Pandas
- NumPy
- Matplotlib
- Seaborn
- Scikit-learn