A fine-grained classification method for tongue images based on fusion of tongue vein features

Abstract

This invention discloses a fine-grained tongue image classification method based on tongue and pulse feature fusion, relating to the field of computer-aided medical diagnosis. The method involves collecting tongue images and corresponding pulse information from multiple subjects, and labeling the segmented tongue regions, tongue image features, and pulse to obtain a training set containing the original tongue image, tongue segmentation labels, tongue image classification labels, and pulse binarization labels. An end-to-end multi-task joint learning architecture is constructed, and a dedicated feature fusion module and fine-grained classification module adapted to tongue and pulse features are designed. After training on the training set, simultaneous and collaborative optimization of accurate tongue segmentation and fine-grained tongue image classification can be achieved. Simultaneously, the method deeply integrates visual features of the tongue image and temporal features of the pulse, fully exploring the inherent pathological correlation between the tongue and pulse, focusing on improving the recognition ability of subtle pathological features such as teeth marks and cracks, comprehensively improving the accuracy of fine-grained tongue image classification, and thus promoting the development of TCM tongue diagnosis towards objectivity, standardization, and precision.

Images