RSL-Alphabet-Recognition

This repository is primarily aimed to use Deep Learning to recognize the alphabets from the Russian Sign Language (RSL).

The core objectives of this mini-project are:

1. Achieving accuracy similar to what Deep Learning has brought to recognition of other sign languages such as the American Sign Language.
2. Handling data scarcity

Dataset

Courtesy of: https://github.com/hse-sl/rsl-alphabet-dataset (20 images/ alphabet)

Results with Transfer Learning

resnet34 - epochs = 15 - max_lr = 0.01 - val_acc: 0.6719

resnet34 - epochs = 15 - max_lr = 0.001 - val_acc: 0.8281

resnet50 - epochs = 15 - max_lr = 0.0025 - val_acc: 0.7203¶

resnet50 - epochs = 30 - max_lr = 0.001 - val_acc: 0.8594

Next Step:

1. Employ Few Shot Learning