Classification of Tongue Images Based on Doublet SVM

Tongue diagnosis is one of the main components of traditional Chinese medicine (TCM). Developing an objective and quantitative recognition model is very importantly and useful in the modernization of TCM. Currently, major problems in digital diagnoses of tongue images are extracting suitable features and building a high-performance classifier. To address these two issues, we present a robust approach to infer the pathological characteristics. In contrast to other methods, this method makes full use of the local information of tongue images and similarities among tongue images. Our method includes the following three steps: (1) we exact HOG features based on theory of local object appearance and shape; (2) the most similar tongue images are found that belongs to the same label and belongs to the different label, which are then used to build a new sample for Doublet; (3) we calculate the distance metric M by the SVM classifier and doublets; and (4) we make prediction. Experimental results show that prediction accuracy of our method is 89.1% and achieves a specificity of 61.3%. Moreover, the Sensitivity is 95.8%. The work is helpful in the area of medical for detection and prevention of diseases.