Multi-focus Image Fusion Based On Consistency-constrained Non-negative Sparse Representation

Multi-focus image fusion is a branch of image fusion.Its main idea is to extract the focus regions in the source images and fuse them into a new image,which contains more clear areas than any source image and is helpful for the understanding of the scene or further image processing.In recent years,due to the enhanced performance of sparse representation?SR?in the image processing field,more and more researchers have applied SR to image fusion.Many SR-based image fusion algorithms have been proposed.Among the existing SR-based image fusion algorithms,the non-negative sparse representation?NNSR?-based fusion algorithms jointly impose the sparsity and non-negative constraint on the representation coefficients,which makes the representation coefficients purely additive and have clearer physical meanings.However,the existing NNSR-based image fusion algorithms still have the following deficiencies:?1?During non-negative sparse coding,each image patch is sparsely coded independently and the similarity among image patches is not taken into account,resulting in the incomplete extraction of essential features from source images;?2?The constructed non-negative dictionary is generally too redundant.Based on these observations,we first construct a consistency-constrainted non-negative sparse representation?CCNSR?model and then apply it to multi-focus image fusion in this dissertation.The experimental results demonstrate that the proposed algorithm outperforms existing multi-focus image fusion algorithms.The main work of this dissertation is as follows:First,a compact non-negative dictionary is constructed for NNSR model.In addition to the non-negative constraint,orthogonal constraint is imposed on the dictionary atoms to reduce the redundancy of the non-negative dictionary.The experiment shows that the constructed nonnegative dictionary has powerful representation ability and can also reduce the computational complexity of the fusion algorithm.Secondly,a CCNSR model is proposed by combining consistency constraint with the traditional non-negative sparse representation model,leading to similar representation coefficients for similar patches under the same non-negative dictionary.Experimental results show that the proposed CCNSR model performs better than the traditional NNSR model in the extraction of useful information from source images.Finally,a CCNSR-based multi-focus image fusion algorithm is proposed.The algorithm mainly includes the following five components:?a?Divide the source images into non-overlap image patches with a fixed-size window and then vectorize these image patches with lexicographic order;?b?Construct a compact non-negative dictionary by using the proposed dictionary learning model,where the training samples are constructed from high-resolution nature images;?c?Sparsely code the source image patches by using the compact non-negative dictionary and the proposed CCNSR model;?d?Construct the pixel-wise label matrix?also called as decision map?by mapping the patch-wise label matrix,which is obtained by using the“₂l-norm choosing-max”fusion rule;?e?Generate the fused image based on the obtained decision map in the spatial domain.Experimental results demonstrate the validity of the proposed fusion method for multi-focus image fusion.