Image Super-resolution Network Based On Multi-scale Feature Fusion

In the case of limitations of hardware devices,super-resolution algorithm is used to improve the resolution of the image and recover the details of image.This thesis aims to discuss the super-resolution algorithms based on the convolutional neural network.In order to solve the problems that the existing super-resolution networks only use attention mechanism in a single-scale space and fail to utilize the multi-scale information of image,two novel super-resolution networks are proposed.The existing super-resolution network based on the attention mechanism only models the channel correlation or long-range spatial dependence in the single-scale space,ignoring the mutual guidance of multi-scale spatial information.To solve this problem,this thesis proposes a super-resolution network based on low-resolution spatial reference recalibration and contrast attention mechanism.The whole network adopts the backward fusion mechanism to fuse the output of each recalibration residual block,from the deep layer to the shallow layer,to make full use of the hierarchical feature.The recalibration residual block is a residual structure composed of the recalibration module and the contrast attention module.The recalibration module encodes multi-scale information,and uses low-resolution space statistics with larger receptive field to recalibrates the original-scale spatial features,which can enhance the high-frequency information adaptively.The contrast attention module adopts contrast pooling that is more suitable for the low-level computer vision task,which can realize the adaptive selection of each channel feature.The proposed super-resolution network is lightweight and efficient while achieving better results.The multi-scale super-resolution networks often adopt the linear method of stacking or adding feature maps to fuse multi-scale features,which cannot realize the adaptive adjustment of the receptive field.In order to realize the adaptive extraction of multi-scale information,this thesis proposes the non-linear perceptual module,and combines the local residual nesting and the global multi-cascade mechanism to build the whole network.The non-linear perceptual module adopts the second-order attention mechanism to conduct the weighted fusion,which overcomes the shortcomings of the existing super-resolution models to fuse multi-scale information in a linear manner,and can realize the adaptive extraction of image multi-scale features.In addition,a residual nesting structure is developed based on non-linear perception modules to make full use of local feature.The whole network adopts a global multi-cascade mechanism to fully extract the non-local hierarchical features from all the convolutional layers.The proposed method can achieve superior performance with moderate parameters.