Cracked Tongue Recognition Based on Deep Features and Multiple-Instance SVM

Yushan Xue, Xiaoqiang Li, Qing Cui, Lu Wang, Pin Wu

Abstract: Cracked tongue can provide valuable diagnostic information for traditional Chinese Medicine doctors. However, due to similar model of real and fake tongue crack, cracked tongue recognition is still challenging. The existing methods make use of handcraft features to classify the cracked tongue which leads to inconstant performance when the length or width of crack is various. In this paper, we pay attention to localized cracked regions of the tongue instead of the whole tongue. We train the Alexnet by using cracked regions and non-cracked regions to extract deep feature of cracked region. At last, cracked tongue recognition is considered as a multiple instance learning problem, and we train a multiple-instance Support Vector Machine (SVM) to make the final decision. Experimental results demonstrate that the proposed method performs better than the method extracting handcraft features.