Research On Accurate And High-Perceptual-Quality Single Image Super-Resolution

Image super-resolution aims to reconstruct a high-quality high-resolution(HR)image from its low-resolution(LR)version.Image super-resolution has been studied for decades,which plays an important role in many fields,such as video surveillance,remote sensing,and image sensing.In this paper,we propose three novel neural networks for single image super-resoluion(SISR).The goals of these methods are to generate accurate HR images,high-perceptual-quality images,and find lightweight networks,respectively.SISR is known to be an ill-posed problem.Lots of tiny textures in HR images are lost in LR images.Making full use of the information in the LR image is the key to generate highquality images.To address this issue,we propose a novel Multi-Scale Residual Hierarchical Dense Network for Single Image Super-Resolution(MS-RHDN),which tries to find the dependencies in multi-level and multi-scale features.Specially,we apply the atrous spatial pyramid pooling,which concatenates multiple atrous convolutions with different dilation rates,and design a residual hierarchical dense structure.The atrous-spatial-pyramid-pooling module is used for learning the relationship of features at multiple scales;while the residual hierarchical dense structure,which consists of several hierarchical dense blocks with skip connections,aims to adaptively detect key information from multi-level features.Meanwhile,dense features from different groups are connected in a dense approach by hierarchical dense blocks,which can adequately extract local multi-level features.MSE-based super-resolution networks usually produce images with over-smoothed and blurry edges and lose a lot of high-frequency details.In the third part of this paper,we use two novel generative adversarial networks(GAN)to produce photo-realistic images.The discriminator of most GAN-based super-resolution networks gives a single score of fake or real on the whole image.However,this manner is coarse.To address this issue,we propose a novel Finegrained Attention Generative Adversarial Network for SISR(FASRGAN).We use a Unet-like network as the discriminator,which has two outputs,a score and a mask.The mask has the same spatial size as HR/SR images,standing for the degree in every position.We combine the mask and MSE loss,making the generator focus on the fake part of SR images.In addition,the generator and discriminator are usually two independent networks.We also propose a novel collaborative network(Co-SRGAN),where the shallow feature extraction part of generator and discriminator are collaborated.In a collaborative manner,the feature extraction is more efficient,which improves the ability of producing high-quality images.The application of super-resolution neural networks in mobile devices is of great significance.The fourth chapter of this paper proposes a new type of lightweight super-resolution network,SFSRN.SFSRN first reduces the spatial size of the feature maps by desubpixel,and then performs deep feature extraction on a small feature map,which greatly reduces the computational complexity of the model.Experimental results show that the proposed method is more effective than other lightweight models.The fourth chapter of this paper uses the real superresolution training data in NTIRE 2019 to train the lightweight network,and deploy it to mobile devices based on the TensorFlow Lite.The quantitative comparisons and visual comparisons on benchmark datasets with stateof-the-art methods illustrate the superiority of our proposed methods.The application of superresolution images in object recognition networks further proves that methods proposed by this paper have powerful reconstruction capabilities and excellent super-resolution effects.