Multi-focus Image Fusion Based On Super-pixel Clustering And Unified Low Rank Representation

Due to the limited field depth of optical imaging system,the objects only at a particular distance from the camera are in focus and can be sharply captured whereas objects as other distances are defocused and blurred.In order to accurately perceiving and analyzing information of a scene,it is desired to obtain images with every object clearly conveyed.Multi-focus image fusion is a simple yet efficient way to solve this problem by integrating multiple images of the same scene captured at different focal points into a single image with all objects in focus.Generally,an object in a multi-focus image is wholly in-focus or out-of-focus as a consequence of the fact that the camera lens usually focuses on an object when we take a picture.That is to say,the object areas have a consistent focusing property.Recently,some new image representation models,such as sparse representation(SR),low rank representation(LRR)and their different extensions,have been employed to multi-focus image fusion.These fusion methods are usually implemented in a patch-based way,which generally ignores such object area consistency during the fusion and easily introduces spatial artifacts in the fused image.Some “jagged” artifacts also exist in the regions between the focused and de-focused regions.To address such problems,a novel multi-focus image fusion method is presented based on super-pixel clustering and a unified LRR(ULRR)model,which can achieve a better performance for fusion.First,the source images are segmented into a few super-pixels of irregular sizes,rather than patches of regular sizes,to diminish the “jagged” artifacts and to preserve the boundaries of objects in the fused image.As well,multiple types of features,including color,edges and textures,are extracted from each super-pixel to improve the robustness of the proposed fusion method.Secondly,a super-pixel clustering-based fusion strategy is employed in the proposed fusion method to reduce the spatial artifacts in the fused images and enhance the object regions consistency.This is achieved by using a proposed ULRR model,which imposes the low-rank constraints on each super-pixel cluster rather than on the whole images.It can enhance the consistency of similar regions and improve the distinction between clusters,which is apparently more reasonable for those images with complicated scenes.Moreover,a Laplacian regularization term is incorporated in the proposed ULRR model to ensure the spatial consistency among the super-pixels with the same cluster.Finally,a measure of focus for each super-pixel is defined to determine the focused and de-focused regions in the source images by jointly using the representation coefficients and the sparse errors derived from the proposed ULRR model.Experimental results demonstrate that the proposed fusion method shows obvious superiorities over some state-of-the-arts in terms of visual and quantitative evaluation.