Regresar

Implementation of a Lightweight CNN for American Sign Language Classification

Abstract:

The American sign language is the most popular and widely-accepted sign language for people with hearing difficulties. Computer vision techniques, such as skeleton recognition, depth recognition, 3D model recognition, or deep learning recognition, have helped to develop better systems for sign language classification and detection. Despite the promising results from baseline research efforts, overfitting problems have been detected when the training and testing accuracy are compared. In this work, we propose to exploit the scaling method on EfficientNet, which is a convolutional neural network architecture, in order to uniformly scale all the dimensions of depth, width, and resolution using a compound coefficient. Our results show that the overfitting problem can be solved by incorporating hyperparameter tunning and dropout as a regularization method. We also have the benefit of transfer learning to reduce the training time by reusing the weights of EfficientNet, pre-trained with the ImageNet dataset. Our results are compared with the benchmark paper, proving that our model generalizes better to unseen instances. In the first section of this study, we introduce American sign language meaning, hand gesture recognition, and its related works in the computer vision field. It allows us to mention transfer learning concepts and Efficient Nets architectures. In the second section, we establish the methodology by choosing the B0 model as the architecture selected to test with a Kaggle dataset. Adjusting hyperparameters, we enter in the third section, in the training and testing phase where overfitting problems were solved with high accuracy, finishing talking with contributions and future works.