Lesion Detection in Wireless Capsule Endoscopy Images Using Texture and Color Features

Distinguishing lesions from normal images quickly is the most challenging work during the review of wireless capsule endoscopy (WCE) videos owing to the large number of images and poor resolution. A novel method based on texture primitive histogram and image block dictionary (IBD) was proposed in this study. Each texture primitive contained 32 dimensions of color information features and 52 dimensions of texture features, which were generated using vector quantization and local binary patterns (LBP) and Leung and Malik (LM) filter bank, respectively. The power of this method was demonstrated by distinguishing 4 kinds of lesions (25 of each kind) from 400 normal images. This method was advantageous over the existing methods, which use the color feature or texture feature alone, with a recall of 93% and a specificity of 92.25%.