Data Augmentation and Deep Learning Modeling Methods on Edge-Device-Based Sign Language Recognition

In this paper, methods of realizing sign language recognition (SLR) on mobile edge devices - AutoML and Transfer Learning - are recommended and compared. Their performance in SLR is compared in this article in terms of model size, speed, and accuracy. To perform a more reliable comparative analysis, we first built and applied data augmentation to obtain a benchmark dataset based on Chinese Sign Language (CSL). With this dataset, models were then trained and deployed to mobile edge devices for real-time testing. It is found that data augmentation can effectively improve the diversity of datasets and improve the robustness of the trained model. The results of the model comparison show that AutoML is more dominant in accuracy, and Transfer Learning is more suitable for low-latency applications. In addition, AutoML excels at classifying macroscopically homogeneous images, such as gestures, despite its low training speed. On the other hand, the training process of Transfer Learning is speedy, but low accuracy remains to be its problem. These conclusions are of guiding significance to the lightweight mobile application implementation of SLR.