SLRNet: Shallow Convolutional Neural Network for Efficient Sign Language Recognition

In recent times, recognizing hand gestures from a sign image has become a vital research field in the areas of computer vision and human-computer interaction. Many researchers have taken advantage of various features and deep learning architectures (DLA) to update the performance of existing sign recognition systems. In small datasets such as ASL, shallow DLAs such as convolutional neural networks (CNN) work well. However, the deeper deep learning architectures are overfitted on the small datasets. To overcome this problem, in this work, two shallow CNN models, SLRNet1 and SLRNet2, are proposed. The proposed models use four convolutional layers to extract the features from the sign image and two fully connected layers to recognize the sign. The first and second models gain 98.94% and 99.14% accuracy, respectively. Both models achieve significant improvements with respect to the present state of the art.Contribution— Two regularized hand gesture recognition models, SLRNet1 and SLRNet2, were developed to work efficiently on small datasets without overfitting and perform fine-tuning to achieve better accuracy.