Image Super-Resolution With Convolution Neural Network

As an important branch in the field of computer vision,single image superresolution(SISR)refers to the process of turning low-resolution image into highresolution image.Convolution neural network(CNN)is one of the major method in SISR.Although larger amount of parameters in CNN can boost its performance,the blindly used of parameters would bring more running-time and memory to the training and testing process,which makes the network inefficient.In this dissertation,we use the convolutional neural network to study the task of single image super-resolution.By designing the networks reasonably,fast computation speed and good performance of image super-resolution are both achieved.The contents and specific research results are summarized as follows:First,we propose an efficient multi-scale residual network(EMSRN)for image super-resolution.Compared with Multi-scale Residual Network(MSRN),EMSRN uses efficient multi-scale residual block(EMSRB)to reduce the network parameters without causing performance loss.Besides,we use the AddStruct to give a weighted summary on the output features of each EMSRB.After using these slight improvements,extensive quantitative and qualitative evaluations on benchmark datasets show that the proposed algorithm outperforms MSRN module in terms of Peak Signal-to-Noise Ratio and Structure Similarity Index.Second,we propose a new network with densely residual connection and interpolation.The super-resolution process is firstly divided into high-frequency image reconstruction and low-frequency image reconstruction,respectively.Then,we combine them at the end of the network.We improved the Fast Super-resolution Convolution Neural Network by adding residual connection and dense connection modules to reconstruct the high-frequency image.The nearest neighbor interpolation algorithm is used to reconstruct the low-frequency image.Transfer learning is used to train the whole model.Compared with the existing methods,the proposed superresolution reconstruction network can reconstruct clear images in multiple data sets.Experimental results show that the peak signal-to-noise ratio and the structural similarity of the reconstructed image are improved.