A two-stage approach for mobile-acquired tongue image with YOLOv5 and LA-UNet

Traditional tongue image segmentation methods are based on standardized image collected on professional tongue image acquisition equipment. The tongue segmentation for non-standard images on mobile terminals are scarce, and the loss of information features seriously affects subsequent analysis and research. Aiming at the problem that tongue image data collected in non-standard environment which contains a large amount of background interference information, this paper proposes a two-stage tongue image segmentation method based on YOLOv5 and light attention UNet (LA-UNet). The output of the coarse segmentation stage of the YOLOv5 network is used as the input of the fine segmentation stage of the LA-UNet network. Through cascade combination, the flexibility of the networks is enhanced. The network at each stage can be independently trained, debugged and output results. It can optimize the model and improve segmentation accuracy. Experimental results show that the mean average precision (mAP) of the YOLOv5 network in mobile tongue image reaches 99.5%, and it can accurately segment the general tongue body area. The cascade network composed of YOLOv5 and LA-UNet achieves the mean pixel accuracy (MPA) of 96.7% and the mean intersection over union (MIOU) of 93.1% in mobile-terminal image segmentation, achieving relatively good results. The results of tongue image segmentation using this method are more accurate than traditional methods. It provides a new idea for the research on mobile-terminal tongue image detection and segmentation, and helps to promote the objectification and modernization of tongue diagnosis.