Resumen
Automated tongue segmentation is a critical component of tongue diagnosis, especially in Traditional Chinese Medicine (TCM), where it has been practiced for thousands of years and is generally considered pain-free and non-invasive. Therefore, a more precise, fast, and robust tongue segmentation system to automatically segment tongue images from its raw format is necessary. Previous algorithms segmented the tongue in different ways, where the results are either inaccurate or time-consuming. Furthermore, none of them developed a dedicated, automatic segmentation system. In this paper, we proposed TongueNet, which is a precise and fast automatic tongue segmentation system. U-net is utilized as the segmentation backbone applying a small-scale image dataset. Besides this, a morphological layer is proposed in the latter stages of the architecture. The proposed system when applied to a tongue image dataset with 1000 images, achieved the highest Pixel Accuracy of 98.45% and consumed 0.267 s per picture on average, which outperformed conventional state-of-the-art tongue segmentation methods in both accuracy and speed. Extensive qualitative and quantitative experiments showed the robustness of the proposed system concerning different positions, poses, and shapes. The results indicate a promising step in achieving a fully automated tongue diagnosis system.