Multi-focus image fusion via coupled dictionary training

Abstract

A novel multi-focus image fusion approach using coupled dictionary training is proposed. It exploits the facts that (i) the patches in example data can be sparsely represented by a couple of over-complete dictionaries related to the focused and blurred categories of images and (ii) merging such representations is better than just selecting the sparsest one in the estimate of the original image. Inspired by these observations, we enforce the similarity of sparse representations between the focused and blurred image patches by jointly training the coupled dictionary, and then fuse these representations to generate an all-in-focus image by a fusion rule. The key characteristics of our approach are bridging the gap between coupled dictionaries, combining plain averaging and "choose-max" as an appropriate fusion rule, and forming a more accurate representation, compared to existing approaches which simply admit sparse representation over one dictionary. Extensive experimental comparisons with state-of-the-art multi-focus image fusion algorithms validate the effectiveness of the proposed approach.