A lightweight design of YOLOv5 with hybrid metaheuristic optimization for hand gesture recognition

Abstract

Hand Gesture Recognition (HGR) has become a vital approach in monitoring patients by their medical professionals to mitigate health risks. To recognize hand gesture signs, advanced deep learning architectures have been widely applied recently. Despite these advancements, balancing accuracy and efficiency remains a major constraint for the current models. Hence, advanced object detection methods, such as the You Only Look Once (YOLO), have been increasingly adopted to bridge this gap. Thus, this work designs a lightweight hand gesture recognition by developing a feature extraction strategy and hybrid metaheuristic optimization for the YOLOv5s network. Initially, it considers key features from RGB, depth, and skeleton hand gesture images, involving the extraction of inter-frame, intra-frame, and finger features. Secondly, the backbone of the YOLOv5s network is updated by the ResNet50 to precisely maintain the tradeoff between accuracy and efficiency through the concise learning of the gesture patterns. It potentially captures various dimensions of finger features, such as direction, shape, and quantity extraction, to improve gesture sign detection. Finally, the proposed model utilizes a novel hybrid metaheuristic algorithm design with a Genetic algorithm and Crow Search Algorithm (GCSA), significantly increasing the speed and improving the quality by selecting the optimal set of hyperparameters. The experimental results on the Praxis hand gesture dataset show the superiority of YOLOv5s.