Image Super-resolution Reconstruction Based On Sparse Representation And Neural Network

Image super-resolution reconstruction algorithm can obtain a high-resolution image from a low-resolution image or multiple ones of the same scene by the means of signal processing.This technology can break through the inherent resolution limitation of existing imaging devices and some interference of external environment,by which can improve the spatial resolution.And it has important application value in remote sensing imaging,medical diagnosis,video surveillance and other fields.As the current research hotspot,the super-resolution reconstruction algorithm based on sparse representation obtain the sparse prior knowledge from the training samples combining the sparse representation theory of the signal to effectively recover the lost details in the degraded process,and improves the resolution of the image.In this paper,on the basic of the super-resolution algorithm based sparse representation,we do some researches including sparse coding with regularization technology,dictionary learning based on neural network and image global error compensation,etc.The main achievements are as follows.The image super-resolution algorithms based on regularization technology are realized.The auto-regressive,non-local similarity and manifold learning regularization terms are introduced into the sparse coding objective function in the sparse coding phase.Among them,the auto-regressive regularization term can guide each image patch to select an appropriate auto-regressive model adaptively to adjust the solution space and improve the recovery of local details.The non-local similarity regularization term can obtain image non-local redundancy to maintain image edge information,and the manifold learning regularization term can obtain structural prior knowledge of image to enhance structural information.Based on the regularization technique,on the one hand,an image super-resolution algorithm based on regularization technique and guided filtering is implemented.It make a error compensation of the reconstructed image using the global error compensation model based on weighted guided filter proposed to further restore detail information of the image.On the other hand,an image super-resolution algorithm based on regularization technique and low-rank matrix recovery is implemented.The different algorithms were used to reconstruct sub-images with different features obtained after the low-rank decomposition to effectively utilize the image features as priori knowledge for guiding reconstruction.Experimental results show that the proposed algorithms recover more detailed information on edges,textures,and structures than the conventional super-resolution reconstruction algorithm based on sparse representation,and the values of peak signal-to-noise ratio and structural similarity are higher than some classical algorithms which are compared.An image super-resolution algorithm based on improved sparse autoencoder is realized.Firstly,in the training set preprocessing stage,the high-and low-resolution image training sets are constructed,respectively,by using high-frequency information of the training samples as the characterization,and then the zero-phase component analysis whitening technique is utilized to decorrelate the formed joint training set to reduce its redundancy.Secondly,a constructed sparse regularization term is added to the cost function of the traditional sparse autoencoder to further strengthen the sparseness constraint on the hidden layer.Thirdly,the improved sparse autoencoder is adopted to achieve unsupervised dictionary learning to improve the accuracy and stability of the dictionary.Finally,the learned dictionary is applied to image super-resolution reconstruction.The algorithm learned a more accurate and robust dictionary and achieved the better reconstruction effect.A medical image super-resolution algorithm based on deep belief network is realized.Firstly,the deep belief network is modified according to the needs of dictionary learning.Then the modified deep belief network is used to learn the dictionary about medical images,and the obtained dictionary is applied to medical super-resolution reconstruction.The algorithm can obtain more accurate image features,and better apply the super-resolution reconstruction technique to medical images to play an auxiliary role for medical diagnosis effectively.