Extended Multi-resolution Local Patterns - A Discriminative Feature Learning Approach for Colonoscopy Image Classification

摘要

We propose a novel local image descriptor called the Extended Multi-resolution Local Patterns, and a discriminative probabilistic framework for learning its parameters together with a multi-class image classifier. Our approach uses training data with image-level labels to learn the features which are discriminative for multi-class colonoscopy image classification. Experiments on a three class (abnormal, normal, uninformative) white-light colonoscopy image dataset with 2800 images show that the proposed feature perform better than popular handdesigned features used in the medical as well as in the computer vision literature for image classification.