Research Of Super-Resolution Image Reconstruction Based On Structural Self-similarity Dictionary

Super-resolution reconstruction of a single image can be seen as an inverse problem of reconstructing a high-resolution image from a low-resolution image using a reasonable model or an efficient algorithm.The technology can break through the limitation of image acquisition equipment and the interference of external conditions,and achieve image quality improvement by means of software analysis,thereby saving image acquisition cost and satisfying people's desire for high quality images in digital image,social security,Face recognition,satellite image and video standard conversion and other aspects have important use value and development space.The learning-based image super-resolution method used in this paper is the most valuable and efficient method for reconstructing super-resolution images in recent years.The learning method is based on constructing a set of low-resolution sample images and corresponding high-resolution sample images,then encoding the reconstructed images under low-resolution sample images,and finally recovering high-resolution images by encoding coefficients.The method overcomes the inaccuracy of the description of the relationship model between images,and has the advantages of accurate reconstruction and robustness to noise.Under the learning-based approach,the dictionary learning method is used as the main line to realize the coding.The research based on structural self-similar dictionary learning,convolutional neural network learning,deformation block feature extraction,etc.is studied.The main research work is as follows:(1)The self-similar features of image structure are studied.There are many similar structural features in each region of the entire image and between different regions,and these similar structural features contain a lot of useful edge or texture information,which is very helpful in image reconstruction.In this paper,a non-local block matching method is used to extract the similarity structure features contained in the image.These feature blocks are composed of high/low resolution with structural self-similarity by matching at the same scaling and different scaling.Image block pairs,and finally these image block pairs constitute a training sample dictionary.(2)A super-resolution algorithm for convolutional neural network based on structural self-similarity is proposed.In this paper,a single image super-resolution algorithm for convolutional neural network model with structural self-similarity features is proposed,which fully utilizes the extraction of image information and solves the problem ofinsufficient feature integration.The method firstly combines the image self sample with the structural similarity feature as the training sample in combination with the non-local block matching method with scale decomposition.Such training samples prevent the problem that the sample is too scattered and the training sample is insufficient.Then,the structural features are fully integrated into the reconstructed image restoration information through the training and learning model of the convolutional neural network,and more prior knowledge is obtained by using the strong learning ability of the convolutional network.Next,the image is reconstructed by the optimal dictionary and the learned model coefficients.In addition,it is enhanced later by an iterative back projection algorithm.(3)A single image super-resolution based on the deformation block feature is proposed.Aiming at the problem of insufficient dictionary sparse performance and insufficient noise robustness in single image super-resolution reconstruction,this paper proposes a single image super-resolution algorithm based on deformed block features.Firstly,the method can extract the similarity features of the image and increase the sample size by establishing the scale gold tower model to expand the search range.Then,the geometric deformation of the sample block enlarges the size of the finite internal dictionary to expand the dictionary space and dictionary reconstruction ability.In addition,the affine transformation in geometric deformation also makes better use of the self-similarity of the image.Finally,group sparse learning can be used to enhance the sparse performance of the dictionary by means of clustering,and the robustness of the sparse noise is also improved.In this paper,the existing bicubic interpolation algorithm,sparse dictionary learning algorithm and deep convolutional neural network algorithm are compared.The super-resolution images reconstructed by the two methods proposed in this paper have more intuitive human visual effects,and the quality evaluation indicators are also greatly improved.In addition,the single image super-resolution method based on the deformation block feature improves the algorithm time efficiency.