Study On Image Super-resolution Reconstruction Based On Multi-manifold Structure

Image is one of the main carriers of visual information.Due to various factors,the resolution of the image is often limited,so super-resolution method is often used to speculate and reconstruct the image details.However,most of the existing methods are end-to-end training,and the complex network model has higher requirements for equipment.In addition,the texture reconstruction of different parts of the image is often related to the category of the part.Therefore,in order to improve the performance of super-resolution reconstruction algorithm,this thesis studies the data set preparation and prior information extraction.The main work and achievements of this thesis are as follows:(1)This thesis proposes a hypothesis that super-pixel segmentation can split an image into multi manifold data structures.Different from the general super segmentation method,this method regards an image as composed of several different categories of pixel blocks,and the pixel blocks in each category are on a flat manifold surface,so the reconstructed texture has similarity.Using super-pixel segmentation to segment the image into different parts,each part is independent,which can greatly increase the number of data sets.At the same time,due to the nature of super-pixel,the pixels in each part have the same category,so that the texture of each image in the data set is more single,which can reduce the demand for the complexity of feature extraction network.According to the category of image blocks,a large number of image blocks can be divided into several sub-manifolds.(2)This thesis proposes using image gradient and geodesic distance to improve SLIC.Experiments on bsds500 data set show that the improved super-pixel method can get better segmentation accuracy and more regular pixel blocks with less number of segmentation than other super-pixel methods while removing discrete points.In particular,the ASA reflecting the purity of pixel class in the image block reaches 0.978.At the same time,the method of calculating the number of segmentation adaptively makes the segmentation results of different images more similar in each index.(3)The category of segmented image blocks is used as a priori to guide image super-resolution reconstruction.Based on the bilinear feature which is the basis of image block classification,the feature probability map of super-pixel block is generated.The feature probability map is used as the prior information of image super-resolution reconstruction,and the channel attention mechanism is added to enhance the effective information.The experimental results show that the PSNR of the proposed method is about 0.1?0.2db higher than that of DBPN in x2 amplfication,which is slightly improved.Although the PSNR is similar to that of RCAN,the number of parameters is reduced by half.