Studies On The Theory And Applications Of Patch Priors In Image Processing

As a powerful and widely used medium of communication, images are turning into the main information source for human to understand the physical world. Image processing and understanding are critical components in contemporary sciences and technologies. However, as a result of the imperfect imaging system, the change of the environment and other factors, images are degraded during the formation, transmission, and recording processes, leading to the decrease of signal-to-noise ratio (SNR) and resolution, distortion, blurring, etc. The phenomenon of the image degradation is seriously affecting the effects of its practical applications, therefore, image analysis and understanding need to eliminate these influences. At present, there are two mainly used ways:one is restoring the degraded image by a pre-treatment, the other one taking into account the deviations caused by the degradation factors when design image understanding algorithms.Noise is one of the main factors causing the degradation of the image, meanwhile, image denoising model can easily be extended to other image restoration models, for example, debluring. Therefore, image denoising is still a hot topic in the field of image processing. Urban change detection plays an important role in different applications such as urban evolotion, urban planning and digital city, etc. Since very high resolution (VHR) images can provide more details, more attentations have been payed on urban change detection using VHR remote sensing images in recent years.Image processing algorithms can be carried out at the "patch" scale, by the utilization of this idea and the image patch priors, this paper aims to improve the signal-to-noise ratio of the degraded image and the change detection accuracy of multi-temporal very high resolution remote sensing images. The main contributions of this dissertation are summarized as the following four parts:The first part focuses on the restoration of hyperspectral images which are polluted by additive white gauss noise. Considering the energy concentration property of principal component analysis (PCA) and the adaptiveness of dictionary learning on each principal component image, a new denoising method for hyperspectral remote sensing images which employs principal component analysis and dictionary learning is proposed. It uses the sparseness prior of image patches in the transformed domain. Both experiments on simulated and real hyperspectral images demonstrate that a better denoising performance is achieved by the proposed method, while well preserving the details, and efficiently restraining the block artifacts.The second part proposes a new dictionary learning model. The smoothness prior assumpts that the image belongs to a smooth function space, and the sparseness prior assumpts that every image patch can be sparsely represented over a set of orthogonal basis, a frame or an over-complete dictionary. Combining with the above two assumptions, we propose that the image patch has a sparse representation over a smooth dictionary, and the second order total generalized variation regularized over-complete dictionary learning model is built, with its application on removing white gaussian noise in the image. Compared with the existing dictionary learning algorithms, the proposed one can effectively control the smoothness of atoms, well preserve structure parts of the image. Numerical experiments demonstrate the validity of the proposed method. The third part focuses on the restoration of images which polluted by multiplicative gamma noise. On the basis of the smoothness prior of images, the self-similarity and the sparseness priors of image patches in the log transformed domain, a similar-patch-group based multiplicative noise removal frame, which simultaneously considers three most important image/patch priors, is proposed. Considering the collaborative roles of higher order singular value decomposition (HOSVD) and total variation (TV) regularization, a novel multiplicative noise removal approach is presented, and the corresponding numerical algorithm is given. Experiments demonstrate the advantages of the proposed approach in removing the multiplicative noise and preserving the details near the edges and in the texture area.The fourth part focuses on change detection of multi-temporal very high resolution remote sensing images. By employing the similarities of image patches in the feature space, a new concept, change field, is proposed. The new concept has the meaningful interpretation in projecting the high dimensional feature space into 3-D change feature space. Based on the displacements and the deviations of similarities between image patches in the spatial domain and the feature domain, respectively, change field effectively measures the complex changes between VHR images at the "patch" scale, and improves the inter-class variability between the changed class and the unchanged class. With the help of the change field and the progressive transductive SVM classifier, a novel very high resolution remote sensing image change detection approach is presented. Experiments demonstrate that the accuracy of some existing change detection algorithms can be improved by using change field, and the newly proposed approach can separate the changed class and the unchanged class effectively.