Transfer Learning-Based Convolution Neural Network Model for Hand Gesture Recognition

Hearing-impaired people use signals made by hand to communicate using sign language. Due to large linguistic variation, it is not easy for an ordinary man to understand it, and hence, computer vision researchers see it as a potential research problem. Numerous approaches have been surveyed and proposed in the literature to cope with the problems of recognition of hand gestures. Selecting and segmenting exact hand shapes in complex backgrounds is a big challenge. Thus, despite tailor-made feature engineering, deep learning is seen as an effective approach to handle this concern. In the proposed work, deep learning approaches based on Convolution Neural Network (CNN) that have already been proven in computer vision are analyzed for Kaggle's American Sign Language (ASL) dataset. Various pre-trained architectures are taken for the experimentation work; the best network is selected and further improved with the optimization technique. AlexNet, GoogleNet, ResNet18, and InceptionV3 are trained, and the effect of variation of optimization techniques, namely Adam, sgdm, and RMSProp is analyzed. Performance parameters such as validation accuracy and testing accuracy are measured to determine the efficiency of the model for each alphabet of sign language. The has been successful in accomplishing the test accuracy of 99.1% for transfer learning of InceptionV3 with a sgdm optimizer.