Deep Learning-Based American Sign Language Recognition System

Designed and developed a complete computer vision pipeline for American Sign Language (ASL) recognition using deep learning and convolutional neural networks (CNNs). Built and compared multiple architectures including custom CNNs, MobileNetV2, ResNet-10, and VGG16-BN across ASL gesture, digit, and alphabet classification tasks. Implemented the full machine learning workflow using Python and PyTorch, including dataset preprocessing, augmentation, training, hyperparameter tuning, evaluation, and visualization. Conducted experiments comparing transfer learning against training from scratch and analyzed the impact of learning rates, batch sizes, dropout, and network depth on model convergence and generalization. Achieved near 99.95% accuracy on full ASL alphabet recognition using deeper residual architectures while identifying challenges such as inter-class gesture similarity, overfitting, and domain mismatch in pretrained models. Evaluated performance using accuracy, precision, recall, F1-score, confusion matrices, and training/validation loss analysis to better understand model behavior and real-world gesture recognition performance.