Generalized Image Prior Based Image Super-resolution Reconstruction

The dramatic growth in the availability of image technology and network data has resulted in a great demand for high-resolution images. However, due to the limitations of various degraded factors, e.g., optical blurring, downsampling and noise, the spatial resolution of the images acquired by the practical imaging system is very low. Image super-resolution technique is a powerful means to effectively improve the spatial resolution of the images with the unchanged sensor technology and the existing imaging systems. The super-resolution technique has significant economic benefits and its wide application in many fields, such as video supervision, medical imaging, remote surveillance and digital television signal conversion. This technique has become one of research hotspots in the image processing field.For the critical problems of image super-resolution reconstruction, this thesis mainly focuses on the estimation of nonlinear mapping relationship, the partition of training examples, the construction of feature subspace, the design of regularizations, and nonlocal means prior in the guidance of the generalized image prior. A deep research on single image super-resolution reconstruction algorithms is made in this paper. The main contributions are as follows:1. A weighted boosting (WB) based super-resolution method is proposed. For each test image patch, a weighted boosting compensation scheme is introduced in the kernel partial least squares (KPLS) regression model. The best score matrices can be adaptively computed and then used to estimate the high-frequency detail of the target high-resolution image patches. The proposed WB method somewhat reduce the time complexity of the traditional KPLS algorithm, which is because KPLS uses the same principal components for all the image patches. In addition, the quality of the resultant images obtained by WB is improved.2. To overcome the low speed problem in the neighbor selection, a clustering and weighted boosting (CWB) based super-resolution method is proposed. First, the texture feature vector of each low-resolution image patch is computed. The large training database is divided into a certain number of different training subsets by exploiting the K-means clustering technique. For each test image patch, several neighbor patches are selected in the corresponding training subset. Second, the nonlinear mapping relationship between low-and high-resolution image patches is learned by using the weighted boosting regression model, which is used to estimate the high-frequency detail. The proposed CWB method has higher time efficiency and better reconstruction results.3. To address the problem that the matching precision between the gradient feature vectors of low-resolution image patches is low in the neighbor embedding relationship, a novel super-resolution method is proposed by using subspace projection and neighbor embedding (SPNE). First, the high-dimensional gradient feature vectors are projected into the kernel principal component analysis (KPCA) feature subspace. The KPCA feature vectors are then projected into the modified locality preserving projection (MLPP) feature subspace. In the MLPP feature subspace, the neighbors of each test low-resolution image patch are selected with higher matching precision. The embedding weights are computed by the similarity measures between feature vectors and proportional factors, which are then used to estimate the high-frequency information from the training high-resolution image patches. The SPNE method effectively improves the super-resolution reconstruction quality.4. Considering the multi-directional property of edges in natural images and the directional group sparsity in its main direction, a novel super-resolution method is presented by exploiting both the directional group sparsity of the image gradients and the directional features in similarity weight estimation. First, the curvelet coefficients of the input image are extracted by the curvelet transform. These curvelet coefficients are then partitioned into sixteen different direction subsets. The corresponding directional feature image can be obtained by using the curvelet inverse transform on each direction subset. A combined total variation (CTV) regularizer is constructed, which can preserve the global and local geometric structures of the image edges. A novel directional nonlocal means (DNLM) regularization term takes pixel values and directional information into account for estimating the reconstruction weights. The proposed method combing CTV and DNLM regularization terms can achieve better results in terms of preserving clearer image edges and recovering richer texture details.5. The existing nonlocal regularization methods heavily depend on the selection of search region in the neighbor selection, which is very time-consuming for an empirical study. In consideration of this, a novel superpixel segmentation and nonlocal means regularization (SNLM) based method is proposed, which can effectively improve the performances of image denoising and image super-resolution. This method first uses the superpixel segmentation technique to partition the input image into various irregular small regions. Then, the search region that contains one or more superpixels is adaptively selected by computing the segmentation labels in a small neighborhood of the target pixel position. The neighbors are searched in the extracted search region. The neighbors with higher matching precision are used to construct the nonlocal means regularization term. Without the empirical study on the width of search window, the proposed method can quickly search the neighbors with higher matching precision, as well as further enhance the quality of the reconstructed images.To sum up, on the basis of image processing, machine learning and pattern recognition, this thesis makes a deep research on three aspects:nonlinear mapping relationship between low-and high-resolution images, matching precision and selection speed of neighbors, and directional feature structures of image edges. Five effective single image super-resolution algorithms have been proposed, which achieve better super-resolution recovery than the existing methods, as well as provide new means to solve the key issues of image super-resolution reconstruction.