Abstract:
Abstract: The existing analysis of sublingual vein often uses convolutional neural network （CNN） image classification methods or image segmentation methods to extract. But there is a problem of low accuracy in extracting details of meridians. Therefore， an improved HRNetV2 high-resolution semantic segmentation network algorithm is proposed to extract sublingual meridians. Adopting a high-resolution HRNetV2 network structure， the outputs of sub network structures from high to low resolution are connected in parallel to form multi-scale fused feature maps with higher spatial accuracy， improving the problem of loss of detailed information in sublingual veins. In addition， AKConv， convolutional kernel with arbitrary sampled shapes and arbitrary number of parameters instead of ordinary convolution can improve the convolution’s adaptability to change pulse structure and reduce the problem of under segmentation. The algorithm is validated through data extraction on the tongue image open platform of Anhui university of Traditional Chinese Medicine （TCM） cloud diagnosis technology， with pixel accuracy（PA）， mean pixel accuracy（mPA）， and mean intersection over union（mIoU） of 95.28%， 92.33%， and 93.42%， respectively， which is superior to the Mask-RCNN model， U-Net models， and HRNetV2 model. The improved HRNetV2 method has high accuracy in segmenting sublingual vein images， providing a new method for further quantitative research on pulse color and shape features.

Key words: Key words:sublingual vein segmentation, high-resolution segmentation, variable kernel convolution, irregular convolution, multi-scale feature fusion