Research And Experiment On Image Super Resolution Reconstruction Based On CNN

Image super-resolution reconstruction refers to the process of rebuilding a high-resolution image from a given low-resolution image by specific algorithms and processing processes.Due to the excellent performance of Convolutional Neural Networks(CNN),the super-resolution reconstruction algorithms based on CNN have achieved remarkable results.However,a large amount of convolution means that a large number of parameters will be generated,consuming a large number of computing costs and memory.To solve this problem,this thesis improves image super-resolution reconstruction algorithms to achieve high-resolution image recovery.The main research content of this thesis is as follows:(1)Research on image super-resolution reconstruction algorithm based on information distillation.To study and analyze the traditional image super-resolution reconstruction algorithm based on CNN,in view of the limitations of too many convolution generation parameters that are difficult to apply on small devices,using information distillation technology,improved feature distillation connection functionally equivalent to channel splitting.And introduce the idea of residual learning to improve the expression ability and performance of the model.Experimental results show that the proposed method effectively improves the quality of reconstructed images.(2)Image super resolution model structure based on fused attention mechanism.Each part of the information in the image is necessary for image reconstruction.Considering that the characteristics generated by the network include different types of cross channel and spatial area information,improved the channel attention(CA)and spatial attention(SA)mechanisms in the distillation module,weighted allocation of features for different channels to suppress redundant and useless information,and to make the model focus more on areas of interest.Aggregate these highlighted features together to obtain more representative features.Effectively fuse these two attention modules to form a super-resolution reconstruction algorithm based on the fused attention mechanism,further improving the expression ability of the model.Finally,in order to evaluate the performance of the improved algorithm,the peak signal-to-noise ratio and structural similarity are used to validate the reconstructed image in the common datasets Set5,Set14,BSDS100 and Urban100.Multiple experiments have shown that this method can effectively improve model performance.