Research on Embedded Tongue Image Segmentation Technology Based on Deep Learning

Abstract

With the continuous development of deep learning technology in the field of medical imaging, objective and intelligent tongue diagnosis in Traditional Chinese Medicine has gradually become a research hotspot. However, due to the complex structure of tongue image features and their high subjectivity, as well as the limited computational power of embedded devices, achieving an efficient, accurate, and low-power tongue diagnosis system still faces many challenges. The goal of this paper is to explore a deep learning-based embedded tongue image segmentation method and to design an efficient low-power segmentation network. The segmentation of tongue images is fundamental to the objectification of tongue diagnosis in Traditional Chinese Medicine, as the accuracy of tongue body segmentation directly affects the identification of tongue color, coating color, and morphological analysis diagnosis in the tongue diagnosis system. Tongue image segmentation involves completely isolating the tongue body region along its edge contour from the image, so that the resulting tongue image contains all the image information of the tongue. During the segmentation process, when the color of the lips or skin around the tongue is similar to that of the tongue body, it increases the difficulty of segmentation, which can lead to low efficiency and accuracy in segmentation results, significantly impacting subsequent tongue image analysis. Therefore, this paper introduces an improved Attention UNet model designed with an attention mechanism, and based on this, it undergoes lightweight processing and optimization of training hyperparameters to achieve precise extraction of the tongue body area, while also considering segmentation performance and device compatibility. Ablation and comparative experiments are conducted on the improved components, followed by a horizontal comparison with currently popular and widely used networks, and finally, it is deployed on a Raspberry Pi hardware platform to verify the effectiveness and advancement of the improvements.