Gesture recognition method based on improved YOLOv5 in complex background
Mingming Chai, Xiaobing Zhang, Dongxia Cheng, Wenchao Tong, Kaixuan Wang · 2024
In view of the problems of low accuracy and difficult recognition of gesture detection algorithms in complex backgrounds. In this paper, a gesture detection and recognition method based on improved YOLOv5 in complex backgrounds is studied. Firstly, to ensure that the network focuses more on effective channel features in complex background images, the SE attention mechanism is introduced into both the main network and the neck network. Subsequently, without significantly increasing computational complexity, the BiFPN module is integrated into the neck network to better facilitate multi-scale feature fusion. Finally, a decoupling head is incorporated into the detection network to resolve conflicts between classification and regression during the output of variables for gesture images, significantly improving recognition accuracy. To validate the superiority of the model, a comparison was conducted with the YOLOv5s model on a custom-made dataset featuring complex backgrounds. The results demonstrate superior detection performance, especially with a 2% improvement in [email protected], highlighting enhanced accuracy in gesture detection. These advancements contribute to overcoming challenges associated with gesture detection in complex backgrounds, providing a robust methodology for improving accuracy and recognition capabilities.