High-Resolution network design for tongue image segmentation

A high-resolution network was proposed to solve the problems of severe loss of edge information and low segmentation accuracy. Firstly, the input images were processed to construct a multi-scale subnetwork parallel connection structure. Secondly, the attention mechanism was introduced to construct the feature extraction module to enhance the extraction of global information. Then, the low-resolution semantic information and high-resolution feature information were fully fused by multi-scale features. Finally, the boundary information was further extracted by the spatial pyramid structure. Compared with the original HRNet network, the average intersection ratio (MIOU) and pixel accuracy (ACC) of the proposed algorithm are improved by 2.6 and 0.7 percentage points, respectively, after evaluation on the self-built dataset. Experimental results show that the algorithm can effectively improve the segmentation accuracy, reduce the loss of edge information, and fully meet the requirements of tongue diagnostic instrument.