Image Super-Resolution Based On Sparse Coding Networks

Image super-resolution(SR)reconstruction use the software algorithm to improve image quality,which generating a high-resolution(HR)image that using one or more low-resolution(LR)observations from the same scene.It has overcome the inherent defects of high cost that through hardware to obtain HR image and has important significance in improving the image display,face recognition,auxiliary diagnosis.It has been widely used in computer vision,medical image processing,security monitoring,remote sensing image processing,and other fields.Super-resolution reconstruction methods based on dictionary-learning and deep learning are very popular in recent years.This paper further study the super-resolution methods based on the above two methods respectively,and made some improvements and innovations.The main work of dissertation are summarized as follows:1.We design a super-resolution reconstruction algorithm which based on multi-dictionary learning and class-anchored neighborhood regression.Firstly,the Gaussianmixture model clustering algorithm is employed to cluster the low resolution training features and generate the label informations;Then we use the supervised KSVD algorithm to generate the over-complete dictionary and a discriminative-linear classifier by solving the objective function,which including the reconstruction error and classification error;Finally,each input feature block is categorized by the classifier and reconstructed by the corresponding subclass dictionary and class-anchored neighborhood regression.2.For the limitations of ReLU and SGD in the original reconstruction networks,we using the RReLU and NAG method to improve the three-layers end-to-end convolutional neural network.On the benchmark set,experimental results further prove that our improved algorithm compared with other methods has better performance on the subjective visual evaluation and objective evaluation.3.We address this problem based on a deep learning method with residual learning in an end-to-end manner.We first analyze the relationship between the image degradation model and residual unit in mathematical formulation.Then we consider using the cascading residual unit to simulate the reverse process of image degradation.In residual network,the signal can be directly propagated from one unit to any other units in both forward and backward passes when using identity mapping as the skip connections.Based on it,we let the residual network to fit a residual mapping between the HR images and LR images rather than directly learn a desired underlying non-linear mapping,which can get a faster convergence speed and gain higher resolution accuracy from considerably increased depth.