A lightweight architecture for hand gesture recognition

This paper proposes a lightweight architecture to recognize hand gestures that can be implemented in the resource-constrained device. There are two main components in our proposed architecture. The first component uses segmentation algorithms as preprocessing to remove noise and irrelevant parts from the input data, while the second component employs a classification algorithm to recognize hand gestures. Different lightweight segmentation and classification algorithms were also investigated and customized. Experimental results showed that the proposed lightweight architecture obtained high accuracy with various datasets even with noisy and complicated-background samples, especially with the combinations of DeepLabV3+ as the segmentation method and MobileNetV2 or EfficientNetB0 as the classification method. In addition, the inference speed of our lightweight system can achieve approximately 20 milliseconds with the fastest backbone even without using a high-end GPU.