Research On Single Image Super-Resolution Reconstruction Algorithm

With the rapid development of science and technology,the demand for high-resolution images in various fields of computer vision is also increasing.However,limited by the hardware conditions of the imaging system,it is generally impossible to obtain high-resolution images directly.At the same time,it is very expensive to upgrade the image resolution by upgrading hardware equipment.Therefore,in order to improve the image resolution without changing the hardware conditions,research on image super-resolution reconstruction algorithms has become a hot spot.In this paper,single image super-resolution reconstruction algorithm is studied to improve the quality and efficiency of the image super-resolution reconstruction.The main research contents of this article are as follows:(1)The image super-resolution reconstruction algorithm based on sparse representation is studied,and the dictionary learning method is introduced.The influence of the dictionary size on the results of image super-resolution reconstruction and the complexity of the reconstruction algorithm is analyzed,and the simulation experiment results are given.(2)A fast image super-resolution reconstruction algorithm based on semi-coupled dictionary is proposed.The difference between the high and low resolution sparse representation coefficients is considered.When sparsely decomposing the image blocks,the mapping matrix between the high and low resolution sparse representation coefficients is used to improve the accuracy of the sparse representation.At the same time,nearest neighbor of each atom in high resolution dictionary is found to calculate the corresponding projection matrix which improves the accuracy of the projection matrix.Experimental results show the effectiveness of the algorithm.(3)Image super-resolution reconstruction algorithm based on convolutional neural network is studied.Increasing the size of the convolution kernel can expand the receptive field,thereby improving the reconstruction effect,but at the same time increases the amount of parameters and consumes computing resources.In this paper,dilated convolution is used to increase the receptive field without increasing the number of parameters.In order to avoid the "grid effect" problem of multi-layer expansion convolution with the same expansion coefficient,the expansion coefficient of the three-layer convolution layer in each residual unit in the network model proposed in this paper is set to a zigzag structure.Experimental results show the effect of the expansion coefficient on the reconstruction effect.(4)An image super-resolution reconstruction algorithm based on deep recursive residual network is proposed.Increasing the depth of the network can improve the reconstruction effect,but also brings about network training problems such as network degradation and gradient disappearance.Residual learning can solve these problems well.Therefore,the network model proposed in this paper is composed of two parallel subnetworks.Different sub-networks can extract image features at different levels by using the combination of global residual learning and multi-path local residual learning.And hybrid convolution is used to improve the receptive field of the convolution layer and obtains more image features.Experimental results show the effectiveness of this algorithm.