Research On Sparse Representation For Image Fusion And Image Restoration

Sparse representation is a novel image representation theory, which can representate an image concisely. Based on a specific dictionary, any image can be expressed as a linear combination of a few atoms, which can reveal the intrinsic properties of image effectively. Currently, sparse representation has drawn a lot of attentions from the international and domestic researchers, and certain research achievements have been obtained. However, the theory of sparse representation and its applications are imperfect, and some difficulties need to be studied further. Therefore, this thesis mainly investigates the theory of sparse representation and its applications for image fusion and image restoration.The main contributions of this thesis are as follows:1. A multimodal image fusion method based on joint sparse model is proposed. Firstly, the multimodal images for the same scene form a signal ensemble due to the relationship of different sensors. Then, all signals in this ensemble are jointly sparsely represented as common and innovation sparse components. At last, the fused image is generated from the common and innovation sparse components. Specially, the common and innovation sparse components indicate the intrinsic relationship among the multimodal images, which solve the problem of the complementary information separation effectively.2. A remote sensing image fusion method based on sparse reconstruction is proposed. The current method based on sparse representation can not construct the dictionary for high resolution multispectral images (MS) effectively due to the shortage of high resolution MS, which hinders the applications and developments of such method. To solve this problem, the proposed method designs a joint dictionaries learning strategy. In this strategy, the dicitionaries for panchromatic image (PAN) and low resolution MS are learned from training set jointly, and the dictionary for high resolution MS is constructed from the dictionaries for PAN and low resolution MS. The proposed method does not need the high resolution MS training set, which makes the method more practical. In addition, the learned dictionaries can reduce the dimensionality of dictionary, speed up the sparse decomposition, and improve the robustness.3. A texture constrained sparse representation model is proposed. In this model, different texture dictionaries are used to sparsely represent the corresponding texture regions. So it can solve the problem that the single dictionary could not represent all texture information effectively. In addition, this thesis proposes an image super resolution method based on the texture constrained sparse representation model. The sparse representation with corresponding texture dictionaries can improve the performance of super resolution in restorating the texture details.4. A simultaneous image fusion and super resolution method based on sparse representation is proposed. The tradition approaches generate a high-resolution fused image by performing image fusion and super resolution separately, which results in the propagation and magnification of artifacts. Noting that the image fusion and super-resolution have some same foundations, the proposed method makes image fusion and super resolution as an organic whole by sparse representation, which can perform image fusion and super resolution simultaneously, and consequently resolve the problem of low resolution image fusion effectively.5. A non-convex group sparse reconstruction model and the DL-GSGR dictionary learning algorithm for group sparse representation are proposed. Group sparse representation employs the sparse prior information and the intrinsic structure of sparse signal, which surpasses the traditional sparse representation. For the group sparse reconstruction problem, this thesis develops a non-convex group sparse reconstruction model. In this model, the non-convex (?)2,p(0<p<1) norm is used to measure the group sparsity and describe the group structure. The uniqueness and robustness of this model are investigated. Then, a re-weighted (?)2,1norm optimizatipn metod is designed to solve the proposed non-convex model, which can improve the performance of group sparse reconstruction algorithm. For the dictionary construction, this thesis proposes the DL-GSGR which applies the graph regularization to preserve the geometrical structure of atom, and improves the ability of learned dictionay. The non-convex group sparse reconstruction model and the DL-GSGR improve the theory of group sparse representation. In addition, this thesis proposes a3-D medical image denoising algorithm based on the group sparse representation with DL-GSGR. In this denoising algorithm, a3-D processing mechanism is designed which can utilize the correlations among nearby slices effective, and a temporal regularization can preserve the relationship among the nearby slices.