Researches On Sparse Representation For Image Restoration And Recognition

Sparse representation is an effective representation theory, which simulates the sparse coding mechanism of cortex in the mammal and can represent the signal as a linear combination of a few atoms selected from the dictionary. The selected atoms and their corresponding sparse coefficients can reflect the intrinsic features and inherent structures of the signal. Recently, the sparse representation and its applications have attracted extensive attention of researchers and become one of the hottest and most difficult subjects in many fields, such as image processing, computer vision, pattern recognition, and so on. This dissertation will mainly research on two core problems of the sparse representation: sparse decomposition and dictionary updating, as well as their applications to image restoration and recognition.The maim contributions of this dissertation are listed as follows: 1. Random refined sparse decomposition for the image denoisingThe traditional sparse decomposition based image denoising methods often pursuit the sparest coefficients for approximating the no-noisy image. However, solving the sparest coefficient is a NP-hard(Non-deterministic Polynomial-time Hard) problem, and difficult to obtain the accurate solution. To address this issue, a MMSE(Minimum Mean Squared Error) estimation based sparse decomposition algorithm has been proposed, which can randomly obtain several sparse coefficients to approximate the MMSE, and create a better estimation for the no-noisy image. In addition, a multiple atom selection strategy has been adapted to reduce the iterations of the proposed algorithm and a FDR(False Discovery Rate) has been used to refine the randomly generated sparse coefficients, thus improving the reconstruction accuracy. Compared to the traditional orthogonal matching pursuit algorithm, the proposed method can has a 10% improvement on the reconstruction accuracy and 4 times less iteration numbers. 2. Structured sparse dictionary construction for the medical image denoisingThe traditional dictionary construction method only learn one over-complete dictionary for image representation. The atoms of the over-complete dictionary has a high redundancy, thus will decrease the accuracy and speed of the sparse decomposition. Meanwhile, a single dictionary is hard to effectively represent the complex structural information in the image.To solve the problem of the high redundancy in the over-complete dictionary, a clustering strategy has been used to divide the dictionary into several sub-dictionaries. Then, for the 3-D medical image denoising, a 3D joint structural dictionary construction method has been proposed, which can simultaneously utilize the correlations within and among slices, thus improving the denoising performance. Compared with the 2D K-SVD dictionary construction algorithm, the proposed method has a 3 d B improvement(in Peak Signal Noise Ratio, PSNR) on the 3D medical image denoising.For the problem that a single dictionary has the poor ability for the representation, a multi-scale structural dictionary construction algorithm has been proposed. It can train different structural dictionaries for different structures and incorporate a multi-scale geometric mechanism into the dictionary learning, thus enhancing the capacity for the representation of small details. In addition, for the medical OCT image denoising, by considering the high correlations among OCT slices, this dissertation propose to learn multi-scale structural dictionary from the nearby high-resolution slices, thereby improving the quality of the learned dictionary. Compared with the state-of-the-art denoising method-BM3 D, the proposed approach can has a 0.42 d B improvement(in PSNR) on the OCT image denoising. 3. Semi-coupled double sparse dictionary construction for the simultaneous denoising and super-resolution of the OCT imageDue to the limitation on the illumination intensity and speed for OCT imaging, low-SNR(signal to noise ratio) and low-resolution image can usually be obtained. To get the high-SNR and high-resolution OCT image, the common way need to separately do the denoising and super-resolution, which will easily amplify the denoising artifacts in the super-resolution stage. To address this issue, this dissertation first acquires a large amount of matched high-SNR, high-resolution and low-SNR, low-resolution training data, and learns the semi-coupled double sparse dictionary together with the mapping function for sparse coefficients. The trained double dictionary and mapping function can effectively reflect the relationship between the low-SNR, low-resolution and high-SNR, high-resolution image spaces, and thus achieves the simultaneous denoising and super-resolution of the OCT image. Compared with the original OCT imaging technique from the Bioptigen Company, the proposed method has 4 times less acquisition time and effectively removes the noise of the OCT image. 4. 3D adaptive sparse decomposition for the OCT image compressionStorage and transmission of the high spatial-temporal resolution OCT data consumes a large amount of memory and communication bandwidth, and creates great challenges for the storage of the clinical data and remote diagnosis. The current sparse representation based image compression algorithm only designs for the 2D image and does not consider the high correlations among the slices in the 3D OCT image, thus limiting the compression performance. To further improve the performance of the sparse representation for the OCT image, a 3D adaptive sparse decomposition algorithm has been proposed. It cannot only utilize the correlations among nearby slices in the OCT image, but also reflect the differences among nearby slices to well preserve the different structural information. In addition, a 3D adaptive sparse coefficients encoding strategy has been designed to enhance the effectiveness of the compression and restoration. Compared with the traditional JPEG2000 and MPEG4 compression methods, the proposed method can has an averaged 5 d B improvement(in PSNR) for the compression ratio ranging from 10 to 40. 5. Multi-scale adaptive sparse decomposition and discriminative sparse dictionary construction for hyperspectral image land-cover recognitionIn the hyperspectral land-cover recognition, the current sparse representation based method only exploits the spatial-spectral information within a fixed-scale(size) window for the recognition. However, a single window scale is hardly adapted for the complex spatial structural information in the hyperspectral image. To address this issue, this dissertation adopts a multi-scale window and proposes a multi-scale adaptive sparse decomposition algorithm. This algorithm cannot only exploit the correlations among different scales, but also consider the differences among scales. Compared with the single-scale sparse representation based method, the proposed method has a 4.2% improvement on the overall recognition accuracy.Most of the sparse dictionary construction methods focus on the representative ability and ignore its discriminative capacity. For the hyperspectral image recognition problem, an effective discriminative dictionary learning algorithm has been proposed. This method sufficiently utilizes the class information of the training set and dictionary atoms to accelerate the training process and increase the discriminative capacity for the dictionary. In addition, by combining with the discriminative dictionary learning algorithm, a superpixel-based discriminative sparse model has been designed. Compared with the traditional single-pixel based recognition method, the proposed method can simultaneously identify the superpixel consisting of multiple pixels with similar spectral characteristics. Compared with the traditional single-pixel based sparse representation method, the proposed approach has a 3.8% improvement on the overall recognition accuracy, while still requiring 10 times less of recognition time. 6. Multi-task adaptive sparse decomposition for Gabor-feature based face recognitionGabor transform can extract the important orientation and scale feature of the human face and so has been widely applied for the face recognition. The traditional Gabor feature based sparse representation recognition method simply concatenates the Gabor features from different orientations and scales into a single column and can hardly exploit the information among them. For the effective utilization of the Gabor features, this dissertation considers the recognition of features from different orientations and scales as different tasks, and proposes the multi-task adaptive sparse decomposition algorithm. This method can exploit the complementary yet correlated information among different features. In addition, this dissertation also designs a Gabor feature local region based recognition method, and proposes a local region adaptive fusion strategy, which considers the structural characteristics of the face and external inferences. Compared with the traditional sparse representation based face recognition method, the proposed method can has an average 15% improvement on the recognition accuracy for the occlusion and noise corruption conditions.