Research On Single Image Super-resolution Reconstruction Technology Based On Deep Learning

It is usually the case that the captured image often degrades and cannot be used due to the limitation of image acquisition device and the collection environment.And the image restoration technology mainly focuses on how to restore images with reduced quality through algorithms.The image restoration techniques mainly consist of super-resolution reconstruction,image denoising,image de-raining,image de-fogging and image inpainting.Super-resolution reconstruction is a technique concentrate on performing a high-resolution image with a down-sampled low-resolution one by utilizing some algorithms.As one of the core technologies of image restoration,super-resolution reconstruction is often used as a pre-order step for several advanced visual tasks such as object detection and image classification.And in our daily life,super-resolution reconstruction is widely used in the fields of satellite imagery,medical imaging,and surveillance security.Since essence of super-resolution reconstruction is to magnify the low-resolution image and perform additional information completion on it,this is a typical ill-posed inversion problem.With the development of deep learning technology in recent years,the current mainstream super-resolution reconstruction method outputs a higher super-resolution image by constructing a deep network and training the model with enormous of samples.However,most of these methods have the problem of network architecture redundancy and low feature utilization within the network.In this paper,we explore the single image super-resolution reconstruction task from the aspects of model compression and improving the reusability of convolutional features.And all of these models proposed in this paper have reached the state-of-the-art performance.We firstly proposed a multi-path network architecture with dense connections that utilized the convolution features in the model more efficiently through more complex network structures.The model can take advantage of the input a priori information and consider more contextual information when restoring high resolution images by combining feature maps from different levels.Extensive experiments demonstrate that the model can reach the state-of-the-art performance with smaller model storage space.Then,we suggest a fully convolutional network based on recursive architecture from the perspective of model compression.While in the recursive model,we introduce the concept of local residuals and global residuals.The local residual refers to the residual for the initial feature learned by the recursive module,and the global residual refers to the residual for the network input reconstructed by the last convolutional layer.In the recursive module,a triple-path convolutional connection is used for feature mapping.The introduction of a 1×1 convolution in the triple-path convolutional block not only compresses the input characteristics,but also reduces the amount of model parameters.In addition,weight sharing mechanism is introduced in the recursive structure to compress the model parameters.Experiments show that the model can achieve the sate-of-the-art reconstruction results with extremely few parameters.Finally,we analyze and summarize the commonalities and limitations of existing deep learning-based image super-resolution reconstruction models,and propose a Squeeze and Excitation Network.The model assumes that convolution features from different depths or convolutional features from different channels have different contributions to the high-resolution reconstruction layer.We introduce an evaluation network to estimate the importance of the convolutional feature maps and adjust them according to the scoring result.Besides,densely connected structure can greatly multiplex these convolutional features and at the same time,ensure that the network can be trained properly.The model simultaneously feeds the convolution features from different depths to the residual reconstruction layer,and enhances the association between the residual module and the layers in the network.Experiments illustrate that,the model can not only achieve the state-of-the-art performance on PSNR and SSIM,but also has higher robustness to complex texture images.