Automatic Classification Based on Features Fusion for Upper Gastrointestinal WCE Images

Wireless Capsule Endoscopy (WCE) is an important clinical application which suffers a time-consuming review procedure. A WCE images automatic classification algorithm which fuses color and texture features is proposed to alleviate the burdensome task. A pre-process that aims to label shadow and highlight is implemented via a new automatic tuning algorithm based on superpixel-level. Hue-Saturation (HS) histograms and a new texture feature named colour scale invariant local ternary pattern (CSILTP) are extracted as local descriptors. A strategy of feature fusion that utilizes discrimination power analysis (DPA) is applied to reduce the dimension of features. The random forest (RF) classifier is then used to discriminate WCE images. The experiment results indicate a better performance of proposed method compared with some existing approaches.