DSE-Net: Deep Semantic Enhanced Network for Mobile Tongue Image Segmentation

Tongue diagnosis plays an important role in traditional Chinese medicine (TCM) because of noninvasive for health assessment. Taking advantage of the portability of mobile devices to develop a tongue diagnosis system has aroused widespread concern in artificial intelligence community. However, mobile tongue image segmentation is challenging on account of low-quality image and limited computing power. In this paper, we propose a deep semantic enhanced (DSE) network to address these issues. DSE-Net consists of a lightweight feature extraction module, efficient deep semantic enhanced module and the decoder. The encoder adopts shufflenetv2 units as backbone for the reduction of computing pressure from mobile devices and the DSE module with multi-scale feature aggregation is designed to improve the network’s recognition of tongue position. In addition, the decoder is designed not only to recover semantics, but also to embed global features from the shallow network for further boosting the retrieval performance. Extensive experiments are conducted on two diverse tongue image benchmarks, including the public tongue dataset collected by special image acquisition device and the lab-made dataset gathered by mobile phones in various uncontrolled environments. The experimental results show the proposed method’s efficiency and accuracy which outperforms the state-of-the-art methods for mobile tongue image segmentation.