Ensemble patch sparse coding: A feature learning method for classification of images with ambiguous edges

Abstract

Sparse coding methods with a learned dictionary have been successful in several image classification problems. However, sparse representations from a unit dictionary may not contain full information when images are affected by environmental factors such as light, shadow, background, and so forth. In addition, sparse features formed by one dictionary can fall into a trap of singularity in training. To handle these problems, especially in images with ambiguous edges, we propose a new sparse coding method based on an ensemble of image patches.The proposed method includes an sparse-coding based image classification framework using image patches and their effective ensemble in an attempt to extract inherent structures from ambiguous-edge images. First, we transform such images into overlapped patches for better classification performance. Then, we assign patch-wise weights and seek to obtain optimal weights not by a single sparse representation but by ensemble learning. For obtaining optimal weights, we propose a two-step update scheme. We collectively update the weights of all patches in misclassified images first and then propagate the weights of misclassified patches to those of other overlapping patches in the images. Experimental results on the Northeastern University surface defect dataset and a close-up skin dataset show the proposed method achieved better classification accuracy than some existing methods and demonstrate the potential advantage of the proposed method in ambiguous-edge image classification.