Purpose/Significance To apply deep learning technology to achieve the purpose of tongue image analysis automation, so as to provide references for the standardization of tongue image of traditional Chinese medicine (TCM) , and further promote the modernization of TCM diagnosis and treatment technology. Method/Process It develops a new semantic segmentation loss function with region-based correlation and label relaxation to enhance the capability of tongue image segmentation model to learn pixel relationships and handle mislabeled data. Additionally, leveraging inherent color-related priors in tongue image features, the model is simplified by decomposing it into two multi-label classification tasks, thereby accelerating model training and reducing its complexity. Result/Conclusion The proposed algorithm is proven effective on a self-constructed dataset, attaining a high 96.57% MIoU in tongue segmentation, and demonstrating strong performance with a macro F1-score of 88.58% and average accuracy of 82.59%.