Research Of Pixel-level Image Fusion Based On Sparse Representation

Subject to the nature of different sensors,sensor information obtained by different modalities are not similar.There are redundancy and complementary information of these sensors.Even one sensor with different transducer parameters or focal length settings can also obtain complementary information from different objects or backgrounds.Image fusion technology integrated information from different sensors or sensor with different parameters to enhance the fused images,which can improve the reliability of visual effects of human and computer processing.Today,image fusion technology is widely used in light-field camera,remote sensing,mapping,navigation,guidance,medical diagnostics and other fields.In recent years,along with the gradual rise of the sparse representation theory,sparse representation based image fusion has a great development.Sparse representation technology can effectively decompose an image to a set of linear sparse coefficients and an over-complete dictionary.The details of images can be remained perfectly by sparse representation.By refering existing researches all around the world,we conduct in-depth researches on sparse representation based image fusion in this thesis.In order to avoid the coefficients incorrect choise problem,a cartoon-texture based image fusion framework for sparse representation based image fusion is proposed.In this thesis,the SKR feature is used to measure the distance between pixels.This distance is used for clustering pixels by local-density-peaks clustering method.PCA analysis is used to get the PCA bases of each cluster.The PCA bases are used as a sub-dictionary to construct a compact and informative dictionary.A cartoon-texture decomposition based fusion framework is proposed.We decomposed the source images by cartoon-texture decomposition.For the cartoon components,weighted average method is implemented to remain the low-frequency information.For the texture components,the previous trained dictionary and Max-L1 method is conducted to get more detailed fusion results.The experiment shows the advancement of our proposed method comparing to mainstream fusion methods.To solve the sparse representation based multi-resolution fusion problem,a novel associated coupled-dictionary learning method for multi-resolution image fusion is proposed,in this thesis.In the practice,as the features of each kind of sensors are different,the source images are always with different resolution.In this thesis,we trained two dictionaries with different resolution for multi-resolution image sparse representation.The multi-resolution images are sparse coded to sparse coefficients with the same dimension.By fusing these sparse coefficients and reconstructed the fused coefficients by the high-resolution dictionary,the fused image can be got.In the fusion process,source images are decomposed by gaussian filter to low-fequency components and high-frequency components,which are fused with different fusion rule.In this case,more detailed images can be got.The efficiency of the proposed method is proved by experiment.In order to improve the performance of multi-resolution image fusion,a multi-resolution image fusion method based on low-rank and sparse decomposition is proposed.In the dictionary learning processes,two groups of coupled-dictionary are trained by the sparse and lowrank components of training samples,respectively.A dictionary-optimization method is proposed for low-rank dictionary couple and sparse dictionary couple.Using the optimized low-rank dictionary couple and sparse dictionary couple can get better description of the low-rank components and sparse components of the source images,respectively.In the fusion processes,the source images are decomposed to low-rank and sparse components.The low-rank and sparse components of are sparse coded by the optimised low-rank and sparse dictionary couple,and the coded sparse coefficients are fused by different fusion rule.Experients show our proposed method achieve great performance.This paper presents an image patches geometric-classification based dictionary learning method to get a informative and compact dictionary.In this algorithm,source images are divided into image patches and these image patches are classified according to the geometric properties.Image patches are classified to smooth patches,stochastic patches,dominant orientation patches.For each class of the image block,a sub-ditionary is trained to guarantee the informative of the trained dictionary.Additionally,PCA analysis is implemented to obtain PCA bases of each image patch class.The PCA basis of each image patch class are combined into a dictionary.Since the PCA bases reduce the redundant information,the trained dictionary can be compact.The experiment shows,that using the dictionary trained by our proposed method can not only reuduce the computation cost,but also improve the fusion performance.