Image Super-resolution Based On Deep Feature Discrimination Enhancement

Image Super-Resolution(SR)aims to reconstruct the corresponding highresolution image from one or more low-resolution observation images,and can compensate for the image resolution degradation caused by the imaging process.It has broad application prospects in security monitoring,medical diagnosis,remote sensing,and other fields.Since the low-resolution image lacks a lot of information,there are usually multiple high-resolution images corresponding to it.Therefore,the image super-resolution task is an ill-posed and inverse problem,and how to recover the highresolution image with high quality is a very challenging task.Deep learning technology has good potential in improving the performance of image super-resolution,and has received widespread attention at present.The deep multi-scale structure can express the features of different perceptual domains in parallel,which can effectively reduce the calculation and help alleviate the gradient disappearance problem of the deep network.At the same time,the attention mechanism can promote the adaptively discriminative extraction of deep features.Therefore,this article first focuses on how to use the attention mechanism to effectively process multiscale features,to enhance the reconstruction ability of the network.In addition,in view of the fact that the top-down feedback connection mechanism is conducive to enriching low-level feature representations,thereby enhancing the modeling capabilities of deep network,this article further focuses on how to introduce the attention strategy to design a feedback mechanism from high-dimensional space to low-dimensional space,to gradually improve the quality of reconstruction.The main works include:1)A single image super-resolution reconstruction algorithm based on attention augmented multi-scale network is proposed.This method aims to enhance the discriminative ability and multi-scale representation ability of the network by discriminatively fusing features of different scales.Specifically,the designed method first calculates the statistics of different scale features through a global average pooling operation,and integrates these statistics as a guide to learn the optimal weight allocation for subsequent multi-scale feature adjustment and aggregation.At the same time,the two-level fusion technology,including intra-group feature fusion and inter-group feature fusion,is applied to make full use of hierarchical features and further improve reconstruction performance.The experimental results show that the proposed method exhibits great competitiveness in terms of balancing reconstruction quality,parameters and efficiency.2)A multi-level feedback network based on spatial feature transform for single image super-resolution is proposed.The motivation of this method includes two aspects.On the one hand,the high-frequency information in low-resolution images is adaptively augmented by using the spatial feature transform technology so that it can be effectively utilized to provide a reference for the image reconstruction.On the other hand,the attention mechanism is introduced to enable the high-level information to provide feedback and optimization for multi-level low-level features,which improves the expressive ability and the discriminative power of network simultaneously.Specifically,the difference between the low-resolution image and its blurred image is first calculated,and the spatial feature transform technology is used to non-linearly fuse the difference image into the original image to obtain high-frequency information enhancement.Then,combine high-level features and low-level features to estimate a set of weights to assist in the discriminative selection of low-level information.At the same time,the highlevel features are mapped back to the low-dimensional space,and the results are superimposed on the adjusted low-level features to make up for the lack of information in the low-dimensional space.Experiments have proved the advantages of this method in subjective and objective performance.