American Sign Language Classification using a Novel CNN Architecture - DivisiveNet

Description

• Four models - AlexNet, ResNet-18, MobileNetV2 and DivisiveNet are trained for ASL classification.
• A new CNN architecture - DivisiveNet is developed.
  • Divides the input image continuously until a single dimensional vector is obtained.
  • The single dimesnional vector is further divided using a series of dense layers until 36 output units are produced.
  • Uses batch normalisation, Leaky ReLU after each convolutions and dense layer.
  • Log softmax is used for classification output.
• All the layers except the final dense layer in the SOTA models are freezed for transfer learning.

Dataset

• ASL dataset from Kaggle: https://www.kaggle.com/datasets/ayuraj/asl-dataset.

Code

• Assignment-2.ipynb - Contains code for training and evaluating AlexNet, ResNet-18, and DivisiveNet.
• Assignment-2-mobilenet.ipynb - Contains code for training and evaluating MobileNetV2.
• Assignment-2-overfit-custom.ipynb - Contains code for training and evaluating DivisiveNet without early stopping.