Stacking of SVMs for abnormal region screening in endoscopic images using multisize patches

Screening for abnormalities through endoscopic images faces many challenges. The location, shape and size of the abnormal regions in the image are unknown and vary across images. It is very difficult to determine the appropriate patch-size to use for generally used patch-based approaches. So we propose to use multi-size patches simultaneously to represent the abnormal regions and construct a stacking (ensemble) in which each support vector machine handles one patch size. Multiple-size patches provide multiple level representation of the image content, which can help improve the abnormal region detection. Experimental results show promising performance of our proposed approach.