Supervised Online Dictionary Learning for Image Separation Using OMP

Abstract

In this paper, we propose a new algorithm to perform single image separation based on online dictionary learning and orthogonal matching pursuit (OMP). This method consists of two separate processes: dictionary training for representing morphologically different components and the separation stage. The training process takes advantage of the prior knowledge of the components by adding component recovery error control penalties. The learned dictionaries have lower coherence with each other and better separation ability, which can benefit the separation process in two ways. Firstly, simple sparse coding methods such as OMP can be used to efficiently obtain superior performance. Secondly, well trained dictionaries can lead to satisfactory separation results even when the components are similar. The dictionaries obtained can also serve as good initial inputs for other models using dictionary learning and sparse representation. Experiments on complex images confirm that better results can be achieved efficiently by our method compared to other state-of-the-art algorithms.