Research On Single Image Super-resolution Reconstruction Based On Attention Mechanism And Multi-scale Residual Network

Single image super-resolution has important applications in many fields.With this technology,we can reduce the broadband requirements of image transmission and improve and the accuracy of remote sensing observation and lesion tissue localization.The convolutional neural network with multi-scale feature extraction structure can obtain a large amount of information from low-resolution input,but the information in different channels and different spaces is treated equally in the process of feature extraction.Different frequency information is mixed together,and there is a large amount of redundant invalid information,which weakens the quality of network output image.In this study,the Parallel Convolution Block Attention Module(PCAM)is constructed by setting the channel attention unit and spatial attention unit in parallel.The Convolution Block Attention Module with Variance Pooling(CAVP)module is constructed by adding the global variance pooling operation to the channel attention unit.Then Convolutional Block Attention Module(CBAM),PCAM and CAVP are embedded into multi-scale residual blocks respectively,and three kinds of attention enhanced multi-scale residual modules are proposed.The attention mechanism is used to recalibrate the feature response to important parts and increase the proportion of obtaining useful information.Comparative experiments show that CBAM module has relatively good modulation effect on multi-scale residual block.This paper also proposes two feature extraction subnets GFENet(Feature Extraction Sub-network Based on Gradually Integrated Features)and OFENet(Feature Extraction Sub-network Based on One Time Feature Fusion),which fuse the feature information on different levels in the model in two ways: step-by-step and one-time.Quantitative and qualitative experiments show that CBAM-enhanced Multi-scale Residual Network Based on One Time Feature Fusion(OCMRN)obtained by the combination of OFENet and CBAM has the best performance.When the up-sampling ratio is ×2 and ×3,the reconstruction effect of OCMRN is better than that of LapSRN,MSRN and EDSR.The performance of OCMRN is close to EDSR on scale factors of×4 and ×8,but its parameters are less than half of EDSR.It shows that the OCMRN proposed in this study achieves the optimal compromise between performance and model complexity.It is effective and feasible.