Distortion Adaptive Image Classification - An Alternative to Barrel-Type Distortion Correction

The endoscopes, utilized in computer aided celiac disease diagnosis are equipped with wide-angle lenses, which introduce significant barrel-type distortions. Previous work is on rectifying the distortions prior to the feature extraction. However, due to new arising inadequacies this often even decreases the classification accuracy. The idea of this paper is based on the fact, that there is a correspondence between the position of a patch in the image and the amount (and orientation) of lens distortions. Therefore, in order to classify an image patch, only a reduced training set of similarly distorted patches is considered. We show that in most cases with the new approach higher classification rates can be achieved compared to traditional distortion corrected and uncorrected image classification.