Cross-domain Tongue Image Segmentation Based on Deep Adversarial Networks and Entropy Minimization

The semantic segmentation of tongue image is a key problem in the development of TCM (Traditional Chinese Medicine) modernization, and there are a lot of research dedicated to the development of tongue segmentation. Although the performance improvement in tongue segmentation with the evolution of deep learning, there are major challenges in generalizing it to the diverse testing domain. As we known, the worse the consistency of cross-domain data distribution between source and target domain is, the lower the performance of model in test domain gets. Existing semantic segmentation methods based on supervised learning are difficult to deal with such problems when it is impossible to re-label the tongue image with poor generalization performance in the target domain. To address this problem, we design a adversarial training framework with regularizing entropy on target domain, aiming to enforce high certainty of model’s prediction on target domain during the trend of domain alignment. Specifically, we pre-trained the tongue image segmentation model with deep supervised method on the source domain. In addition to segmentation task, the segmentation model need to regularize entropy of output on target domain and maximally confuse the discriminator. The discriminator tries to distinguish whether the output of segmentation model from the source domain or the target domain. In this study, two datasets is constructed, and the five-fold cross-validation experiment is performed on it. Experimental results show that the tongue image segmentation performance in the open environment was improved by 21.5% mIOU (59.2% → 80.7%) after domain adaptation. As opposed to the pseudo label learning with different thresholds(0.6, 0.9), the mIOU of proposed method increased by 17%, 16.1%. Moreover, as opposed to MinEnt, the mIOU increased by 6%. The tongue images cross-domain segmentation method proposed in this paper significantly improves the segmentation accuracy in the unlabeled target domain by reducing the influence of the cross-domain discrepancy and enhancing the certainty of model output in target domain.