Fast Tongue Image Extraction Fusing Multi-dimensional Information and Concurrent D-search Algorithm

Tongue image extraction is a primary and fundamental step in the objectification of the tongue diagnoses of traditional Chinese medicine. Aiming at the problems that some methods need to adjust parameters by manual and their accuracy and efficiency are not high, in this paper, a kind of novel concurrent d-search algorithm fusing multi-dimensional information is proposed for the application of tongue image extraction in order to realize automatic, fast and accurate tongue image extraction. In our method, original tongue image is considered as a pixel network. Starting from the seed pixel of central region, according to the intensity, hue and relative position information of the pixels of tongue image, a novel d-search algorithm is utilized to search tongue body region in original tongue image. Because tongue coating is surrounded by tongue substance in most cases, we propose a novel position operator to decide whether current pixel is surrounded by tongue substance in our d-search algorithm so as to extract tongue image with coating accurately. In the same time, to improve the efficiency of our algorithm, the original tongue image is divided into several regions (i.e. sub-images), concurrent techniques are utilized to search different tongue regions separately, so as to get the tongue image templates of different regions quickly, and then they are combined into one template. In addition, morphological operators including closing and opening are applied to the template after combination, so that small holes in the template image can be filled. In the end, an and operation is applied to the ultimate template and original tongue image, in order to get extracted tongue image. In our study, we conduct the comparison experiments of 4 methods including the method of manual extraction, our method, tongue image extraction based on greedy rules and tongue image extraction based on random walk. The comparison experiments are conducted from the aspects of accuracy and efficiency. The tongue images of the 5 typical kinds of colors including light red, light white, red, deep red and purple are utilized to test the 3 methods. Furthermore, a tongue image with thick white coating and a tongue image with thick yellow coating are also utilized to test the 3 methods. These tongue image samples cover all kinds of tongue colors and include tongue images with or without tongue coating. Therefore, the results of our tests are persuasive. In comparative experiments, our method achieves quite good effect in the aspect of accuracy and is superior to the current 2 methods in the aspect of efficiency, which can meet the accuracy and efficiency requirements of the objectification of tongue diagnoses. And we have implemented automatic tongue image extraction without adjusting any parameters by manual.