Research On Image Super-resolution Reconstruction Method Based On Residual Network

The quality of image will be reduced due to factors such as equipment and environment during the process of acquisition and transmission.Image super-resolution reconstruction is a software way to supplement detail information.It can improve the clarity of image without increasing the cost of hardware,being extensively applied in public safety,topographic mapping,medical research and so on.The focus on the paper was to study the image super-resolution reconstruction method based on residual network.The paper first conducted an in-depth study of related theoretical knowledge such as image super-resolution reconstruction model and residual network as well.It laid theoretical foundation for using residual network to deal with the problem of super-resolution reconstruction of low-resolution image.Then different residual network models were proposed for the reconstruction demands for RGB image and depth image.One was the super-resolution reconstruction method of RGB image based on accelerated residual network.The other was the super-resolution reconstruction method of depth image based on side output layer supervised double-flow residual network.After each method was explained,different types of low-resolution degraded image were used to test the reconstruction performance of the method.Experiments show that the two methods can show strong applicability in the reconstruction of low-resolution image under the general degradation model and obtain good results in reconstruction.There are three aspects included in the paper as following.Firstly,by combing the current research status of image super-resolution reconstruction methods at home and abroad,it was found that most methods were designed to deal with the reconstruction of ideal low-resolution image generated by a single down-sampling degradation factor whether it was RGB image reconstruction research or depth image reconstruction research.They did not take into account more general degradation factors such as motion blur and gaussian blur.According to the current research,suitable reconstruction network models were proposed from the aspect of improving the reconstruction effects of more general low-resolution RGB image and more general low-resolution depth image in this paper.Secondly,aiming at the problems of multiple network parameters and high time complexity in RGB image super-resolution reconstruction method based on deep network architecture,a super-resolution reconstruction method of RGB image based on accelerated residual network was proposed,referred to as ARSR.To begin with,constructed a global residual network to reconstruct the high-frequency residual information between the low-resolution image and the high-resolution image which can reduce the deep network transmission process of redundant information and improve reconstruction efficiency.Afterwards,the dimensionality of the extracted low-resolution feature maps was reduced through feature shrinking layer which can stimulate fast mapping with fewer network parameters compared to the comparison method.Then,the dimensionality of the high-resolution feature maps was increased through feature expansion layer and the high-frequency residual can be reconstructed with richer information.Finally,the residual image and the low-resolution image were fused to obtain the reconstructed high-resolution image.The proposed method can obtain better reconstruction quality much faster with fewer network parameters.In addition,using this method to reconstruct low-resolution motion blur RGB image also shows good applicability.Thirdly,there are many smooth areas in depth image.In addition,it lacks texture information and the edge structure is more prominent.For these characteristics,this paper improved on the residual network structure and proposed a super-resolution reconstruction method of depth image based on side output layer supervised double-flow residual network,referred to as Do SSR.The specific improvement had the following three aspects.First of all,it constructed a double-flow residual network to reconstruct high-frequency residual information.The shallow network learned the rough edge structure of depth image.The deep network learned the refined content of the edge structure.It obtained the reconstructed residual image by fusing the two parts.The construction of the double-flow residual network can improve the reconstruction quality of depth image edge.Then,it added side output layers at different depth positions of the network for local supervision which can not only reuse image feature learned by different depth convolutional layers and improve the problem of information loss in the deep network transmission process,but also extract feature maps of different size of receptive field and effectively utilize richer image context information.Finally,it introduced inception module in the feature extraction part and used multiple convolution kernels of different scales to extract feature information at different positions of image.The performance gain was obtained by fusing the feature of different positions through aggregation operation.Moreover,the dimensionality reduction processing inside the inception module was also conducive to reducing the parameter scale of the network and reducing the time complexity.In the paper,this method was applied to reconstruct low-resolution motion blur depth image and low-resolution gaussian blur depth image and both achieved good reconstruction effects.The two image super-resolution reconstruction methods proposed in this paper can effectively deal with the reconstruction problems of RGB image and depth image under the general degradation model and achieve the dual goal of obtaining higher reconstruction quality with fewer network parameters,which contributes to theoretical significance and application value.