Research On Improved Image Super-resolution Method Based On Broad Network

In modern society,images are a major way for people to obtain information from the outside world.High-resolution images contain a wealth of detailed information and colors,and can play a very important role in the fields of medical diagnosis,smart devices,and video surveillance.However,in some cases,due to the limitations of the imaging system and imaging environment,we can only obtain low-resolution images.The lower image resolution will bring great challenges to various image processing systems,resulting in degradation of the performance of the image processing system.Therefore,how to improve the resolution of an image has always been a concern.Image super-resolution reconstruction refers to the use of image processing technology to reconstruct a corresponding high-resolution image from a given low-resolution image.In recent years,image super-resolution methods based on deep learning have achieved high super-resolution reconstruction quality,but the training of deep neural networks requires a lot of computing resources(including memory,CPU computing power,etc.).The broad learning system is a single hidden layer neural network with a flat structure,and its hidden layer contains a large number of neurons.Width learning has been widely used in the field of image processing.Based on the broad learning system,this paper studies the problem of image super-resolution.The main work is as follows:First,a broad learning super-resolution algorithm based on iterative back-projection preprocessing is proposed.In the training phase,the algorithm first uses the iterative back projection method to enhance the low-resolution image after bicubic interpolation,then,the enhanced image is compared with the high-resolution image to obtain the residual image,and finally the broad network is used to establish the enhanced image The mapping relationship between the low-resolution image block and the residual image block.In the reconstruction stage,the low-resolution image enhanced by iterative back-projection is input into the trained broad network to obtain the corresponding residual image,and then the residual image and the enhanced low-resolution image are added to obtain the reconstructed high-resolution image Rate image.Experimental results show that the algorithm can effectively improve the quality of reconstructed images.Secondly,because the internal structure and mode of the image are relatively complex,a single model cannot describe the mapping relationship between low-resolution image blocks and high-resolution image blocks well.To this end,this paper proposes a subspace image super-resolution algorithm based on K-means clustering.First,the algorithm uses the K-means clustering method to cluster low-resolution image blocks,and divides the low-resolution image blocks into K subspaces.Then,for each subspace,the broad neural network is used to establish the mapping relationship between the low-resolution image block and the high-resolution image block.In the reconstruction stage,the subspace to which the low-resolution image block belongs is determined according to the distance between the low-resolution image block and the center of each subspace,and then the broad network model of the subspace is used to obtain the corresponding high-resolution image block.Compared with a single model,the multi-broad network model effectively improves the adaptability of the super-resolution reconstruction process.Experimental results show that the proposed algorithm can generate high-quality images with rich details.Finally,in order to improve the quality of the super-resolution reconstruction of the broad learning network image,a quality enhancement post-processing method based on the non-local self-similarity of the image is proposed.Based on the high-resolution image reconstructed by the broad learning network,this algorithm makes full use of the redundant information in the image to enhance the details of the image and further remove the artifacts.Experimental results show that the proposed algorithm can generate high-resolution images with sharp edges and clear textures.