Efficient spatio-temporal modeling for sign language recognition using CNN and RNN architectures

Computer vision has been identified as one of the solutions to bridge communication barriers between speech-impaired populations and those without impairment as most people are unaware of the sign language used by speech-impaired individuals. Numerous studies have been conducted to address this challenge. However, recognizing word signs, which are usually dynamic and involve more than one frame per sign, remains a challenge. This study used Tanzania Sign Language datasets collected using mobile phone selfie cameras to investigate the performance of deep learning algorithms that capture spatial and temporal relationships features of video frames. The study used CNN-LSTM and CNN-GRU architectures, where CNN-GRU with an ELU activation function is proposed to enhance learning efficiency and performance. The findings indicate that the proposed CNN-GRU model with ELU activation achieved an accuracy of 94%, compared to 93% for the standard CNN-GRU model and CNN-LSTM. In addition, the study evaluated performance of the proposed model in a signer-independent setting, where the results varied significantly across individual signers, with the highest accuracy reaching 66%. These results show that more effort is required to improve signer independence performance, including the challenges of hand dominance by optimizing spatial features.