Dynamic Searching and Classification for Highlight Removal on Endoscopic Image

Abstract

Endoscopic imaging is a common clinical modality to inspect surficial abnormality grew on the internal organs inside human body. Covered by tissue fluid, surface of these anatomies tend to be glossy, showing specular reflections from the illumination source. In this paper, we present a novel method for specular region separation and restoration from only a single image. Distinguishing from segmentation methods using simple threshold, our solution treats the separation of highlight pixels as a binarization problem based upon a supervised learning classification algorithm. Also, we propose a multiscale dynamic image expansion and fusion based method to restore the highlighted region. It takes full advantages of propagating the regions with similar structure features to specular regions. Experimental results on the removal of the endoscopic image with specular reflections demonstrate improved efficiency by the proposed method compared to commonly used techniques.