Research On Image Super-Resolution Reconstruction Based On Nonlinear Sample Learning

Image super-resolution(SR) reconstruction can recover the original high-resolution(HR) image from one or more low-resolution(LR) images. The nonlinear-sample-learning based SR has become a research hotspot in the field of image SR in recent years, which mainly consists of sparse encoding based method and neighbor-embedding-based SR. However, this method can only get the shallow representation of the image because of its linear processing mode, and it is difficult to dig the deep geometrical structure feature of the image. Especially, the construction of the feature space and the selection of the neighborhood in neighbor-embedding-based SR have a great impact on the final results, and the structure of the single manifold is not accurate. Therefore, the nonlinear-sample-learning based SR that using a nonlinear approach to embed or encode the reconstructed image, which has overcome the shortcomings of the traditional method of the linear-sample-learning based SR.In this paper, we mainly do a deep research and analysis on the neighborhood selection, feature extraction, multi manifold learning and large multiple reconstruction under nonlinear frame in the process of SR. Considering the shortcomings of the existing methods, we propose an improved method and idea. The detail works are as follows:1. Representative features embedding based SR. In view of the problem of feature extraction and nearest neighbor selection in the neighborhood embedding method, we improve the process of feature extraction and the limit of the fixed size of the neighborhood. Firstly, stacked sparse auto-encoder is introduced to learn image patch nonlinear deep features in order to find a more accurate neighbor, which makes it more accurately to describe the direction and texture properties of the image itself. Secondly, we construct an adaptive neighborhood constraint function, which is adaptive to the size of the neighborhood, and avoid the introduction of the error in the embedding process. Finally, the low resolution input image is reconstructed by the proposed method, and restore the more accurate high frequency details. The experimental results show that the method has good results in both visual effect and numerical value, and is superior to other methods in comparison.2. Multi-manifolds neighbor embedding for SR. Considering a large number of image patches often contain multiple geometric structures and information, the assumption of a single manifold is not accurate in the field of image SR, so the proposed method employs multi-manifold attribution assumption for image patches. We construct the training set based on the multi-manifold learning via the clustering method firstly, and learn the corresponding manifold of the multiple classes. Then, the kernel method is introduced into the LLE algorithm, using the kernel space distance to calculate the nearest neighbor of the sample, and completing the selection of the neighborhood and the reconstruction of the image in the kernel space simultaneously. The proposed method can not only maintain the nonlinear nature of the data but also can inherit the manifold structure of the data. The experimental results show that the method has achieved good results in the visual effect and the numerical index, and the high frequency details are recovered.3. Non-linear compressive sensing using kernel methods based SR. Compressive sensing theory indicates that sparse or compressible signals can be accurately recovered from a surprisingly small number of measurements. In this method, we first extend a linear sparse encoding model to a nonlinear model to obtain a nonlinear compressed sensing framework, and apply this theory to the problem of SR. Then an analytical formula is solved by the method of nonlinear compression, which avoids the large scale of the iterative process, and the HR images are recovered only via a simple least square algorithm so that reducing the time complexity. Moreover, the reconstructed image with a large manification factor can produce a jagged effect, not a clear outline, and a fuzzy edge. So this method enhances the effect of SR under high magnification factor, which overcomes the problem of the traditional model based algorithm for high magnification factor.