Research On CNN Super Resolution Reconstruction Of Single Image

As the carrier of recording information,image plays an important role in real life.For the image,the resolution determines the quality of the image.The higher the resolution,the better the quality,and the more information it contains,which is beneficial to the subsequent application processing.In the actual image acquisition,due to the influence of hardware cost and technology development,the image resolution obtained does not meet the specific requirements.In order to obtain higher resolution images,people began to use a variety of ways to solve this problem,but the cost of improving the hardware is too high,so they pay more attention to the software method.Convolutional neural network(CNN)has excellent feature processing ability and can effectively learn image feature information,which is beneficial to image super-resolution reconstruction.In this paper,based on convolutional neural network,two methods of binary channels multi-scale residual network(BMRN)and multi-scale cross merge network(MSCM)are designed to realize single image super-resolution reconstruction.The main work is as follows:BMRN consists of feature extraction,nonlinear mapping and reconstruction.Firstly,in the feature extraction,the low resolution image is directly processed to extract the feature information from the image;in the nonlinear mapping,the dual path multi-scale residual network structure composed of multiple independent sub networks is constructed to extract high-frequency information,and residual connection is introduced to ensure the stability of the network.In the final reconstruction,sub-pixel sampling is used to obtain the final reconstructed high-resolution image.Simulation results show that compared with Bicubic,SRCNN and VDSR,the subjective evaluation index of BMRN method has obvious advantages,and the edge features and sharpening effect of the reconstructed image are the best.MSCM is an improvement of BMRN,which is also composed of three parts.In feature extraction,multi-scale convolution is used to process the input low-resolution image;the nonlinear mapping part is composed of five cross merging modules,each of which is cascaded by three residual double branch merging structures to promote the information integration in different branches.In the nonlinear mapping module,dense connection and residual connection are integrated to improve the information transmission and ladder In the reconstruction part,the improved sub-pixel sampling is used to achieve,and the loss of feature details is compensated by combining external residual with global residual.Simulation results show that,compared with BMRN method,MSCM method has better image reconstruction effect,and the objective and subjective evaluation index values are optimal.Figure [42] table [8] reference [91]...