Cascaded CNN for Real-time Tongue Segmentation Based on Key Points Localization

Automated Tongue Diagnosis (ATD) is a growing research field in recent years due to global demand of personal health care. Automated tongue segmentation is the fundamental technology for automated tongue diagnosis. It is significant to find an efficient and accurate tongue segmentation algorithm for mobile and embedded devices. However, researches in this field are inadequate. Existing methods are either not efficient or not accurate enough. Our paper first addresses these problems by presenting a class of efficient cascaded CNN models for tongue detection and segmentation, which is both lightweight and accurate compared with other high-performance algorithms. Furthermore, a new annotation method is proposed to reduce the workload of establishing a tongue segmentation database.