Research On The Recovery And Super-Resolution Method Of Degraded Images

The recovery and super-resolution of degraded images has been a research hotspot in the field of computer vision,which is limited by the weather and acquisition equipment,resulting in foggy and low resolution images,making it difficult to obtain key information directly from the images,which seriously restricts the application of computer vision in the fields of autonomous driving and target recognition.Therefore,appropriate methods to recover and super-resolution degraded images are of wide research and application value.Among them,image dehazing and image superresolution reconstruction techniques have strong representation and wide application areas.In this thesis,we study and improve image dehazing and image super-resolution reconstruction methods based on generative adversarial networks,respectively,and the main works are as follows:(1)Proposed a method of recovering foggy days degraded images.To address the problem that current image dehazing methods generally focus on the whole image and thus cannot handle non-uniform fog well,which leads to color distortion in the recovered image,the grid asymmetric convolution module GACB(Grid Asymmetric Convolution Block)is proposed for feature transformation,and image features are extracted by using different sizes of 1×N and N×1 convolution kernels in parallel.The non-uniform fog of different sizes in the image is given full attention.To further exploit the utility of the GACB module,GACBNet is proposed to enhance the image feature communication between different levels based on the U-Net idea.Meanwhile,a multidiscriminator structure with global discriminator and local discriminator is used to discriminate non-uniform fog to further enhance the capability of handling nonuniform fog.Since the existing methods are affected by the overall hue of the foggy image dataset,the overall hue of the recovered image is dark,and the color loss constraint is added to the loss function to recover the image hue close to the original image.(2)Proposed a method of super-resolution reconstruction method for low-resolution degraded images.Aiming at the current image super-resolution reconstruction methods which generally focus on the whole image without highlighting the foreground information and lacking visual focus,and have problems such as low utilization of shallow features and large number of training parameters,a parallel attention module PAM(Parallel Attention Model)based on gated network is proposed to compute in parallel on the residual branch of Residual Block It proposes a non-uniform joint loss to dynamically adjust the weights of channel attention and spatial attention in cooperation with the gated network,so as to focus on the image foreground information and improve the foreground sharpness of the reconstructed image.To make full use of the shallow features,PAMNet is proposed based on the PAM module and jump connection,which retains the shallow features of the image while fully exploiting the detailed features of the image,and preserves the background features and color features of the image while improving the foreground sharpness of the reconstructed image.(3)Designing ablation experiments and comparison experiments.First,for the recovery method of foggy degraded images proposed in this thesis,ablation experiments are designed for GACB module and multi-discriminator structure respectively,and a comprehensive comparison is made with Grid Dehaze Net,GCANet,FFA-Net and other existing methods on different datasets.Secondly,for the superresolution reconstruction method of low-resolution degraded images proposed in this thesis,we design ablation experiments on PAM module and skip connection,respectively,and make a comprehensive comparison with existing methods such as SRGAN,ESRGAN,RFANet,etc.on different data sets.Finally,for the scenes in which the resolution needs to be improved in foggy degraded images,the image dehazing method and the image super-resolution reconstruction method of this paper are used serially to clarify and super-resolve the foggy degraded images.Compared with the image dehazing method only,the serial use of the two methods not only removes the fog in the images,but also highlights the foreground features more,further improving the clarity and quality of the generated images.