An improved tongue image segmentation algorithm based on Deeplabv3+ framework

Abstract

Tongue image segmentation is the key step of traditional Chinese medicine (TCM) intelligent tongue image analysis. The subsequent tongue image quality analysis is directly affected by the precision of segmentation. Deeplabv3+ network has become an excellent algorithm in the field of tongue images segmentation by virtue of its ability to extract multi-scale information and its codec structure. However, there are unclear edge segmentation of the tongue body and missegmentation of small areas in some tongue images. In view of the above phenomenon, an improved algorithm is proposed. Firstly, the network structure is optimized, so that the ability of the network to extract multi-scale information and low-level information is improved. Secondly, a loss function based on edge information is proposed, which makes the network pay more attention to the separation of tongue edges in the process of training. Finally, the segmentation results are post-processed by using the prior knowledge of tongue image, so as to eliminate the phenomenon of misjudgement. The experimental results show that the algorithm significantly improves the ambiguity of image segmentation, and the MIOU value is still increased to 99.13% when the MIOU value has reached 98.77%.