Research On Single Image Super-resolution Method Based On Deep Residual Model

High-resolution images have the advantages of clear picture and rich colors.Many application scenarios have high requirements on image resolution.However,the imaging process is affected by many factors and often cannot obtain high quality images in the real world.In order to solve this problem,the super-resolution technology can be used to recover the loosed detailed information of the image,and thereby improves the resolution.The development of deep learning has made the super-resolution a great breakthrough,which is a research hotpot at home and abroad in recent years.This thesis focuses on the method of super-resolution reconstruction based deep learning.The main research contents are as follows:1.This thesis introduces the research background and basic knowledge of single image super-resolution firstly,then analyzes the research status of this technology and the application of deep learning in super-resolution.As to further improve the quality of reconstruction and boost the performance against existing algorithms,effective superresolution models based on convolutional neural network and generative adversarial network are proposed.2.Traditional feedforward neural networks tend to loss local information when extracting image features.For this reason,an enhanced two-stage super-resolution residual network is proposed,based on the full consideration of the influence regrading information persistence in the process of network propagation.This model can effectively fuse all features learned by each layer while increasing the depth of the network.The first stage of reconstruction is to obtain hierarchical features via the dense residual unit we constructed,and to improve the integration of information.The second stage of reconstruction is mainly to carry out residual re-learning on the high-frequency information obtained in the first stage,and to reduce the reconstruction error.Between these two phases,model introduces feature scaling and dilated convolution to achieve the dual purpose of reducing information redundancy and increasing the convolutional receptive field.3.Most of the existing super-resolution algorithms use the mean square error loss as the object optimization function.As a result,the reconstructed image texture is blurred,which is difficult to satisfy the subjective visual perception.In order to solve this problem,this thesis proposes a novel reconstruction model based on the framework of generative adversarial network,and highly restores the high-frequency semantic features.The generative sub-network is composed of feature pyramids which contain dense residual blocks,completing the reconstruction task of different scales progressively.Stride convolution and global average pooling are introduced in the discriminator sub-network,which effectively captures the data distribution of the reconstructed image,generated from generative sub-network.These components improve the ability to judge the authenticity of the generated image.Finally,the objective function combines the perceptual loss to focuses on the resolved semantic features that affect the visual perception of image.