Research On Image Dehazing Algorithm Based On Generative Adversarial Network

With the rapid development of the economy and informatization in China,computer vision technology is widely used in various fields,however,massive emissions of industrial and traffic pollutants lead to frequent haze occurrences,and the quality of images captured under haze degrades significantly with serious loss of detailed information,which seriously affects the performance of high-level vision tasks based on clear input data,such as object detection,vehicle recognition,and intelligent driving.It is of important research value to reduce or eliminate haze in images,restore image details,and obtain clear images through image processing techniques.Aiming at the problems of inaccurate estimation of atmospheric scattering model parameters and lack of feature screening for current dehazing algorithms,an image dehazing algorithm based on channel attention and generative adversarial network is proposed combined with the advantages of generative adversarial network in image reconstruction.The proposed algorithm learns the mapping from haze image to clear image directly and overcomes the dependence of the atmospheric scattering model.The generator designs multi-scale residual block(MRBlk)to expand the perceptual field,extract multi-scale features of images,learn features shared by haze images and clear images,and reduce information loss.The efficient channel attention(ECA)module learns the correlation characteristics of adjacent channels,dynamically adjusts feature channel weights,provides priority to effective channels,enhances effective feature propagation,suppresses irrelevant features,and improves the feature screening ability of the dehazing algorithm.Patch GAN is used to evaluate images in chunks,improve image discrimination accuracy,and guide the generator to generate more realistic images.Perceptual loss is proposed to reduce the loss of dehazing image features and retain more detailed image information so that high-quality dehazed image can be achieved.The experimental results show that the proposed algorithm is an effective image dehazing method.Average gradient is improved by 4.15% on average,compared with DCP,AOD-Net,Grid Dehaze Net,and FFA-Net.Most existing studies rely on haze and clear image pairs,but in practice,obtaining a large number of matched image pairs is difficult to achieve due to various factors,and the network trained on synthetic datasets does not work well when processing real haze images,leading to the domain shift problem.To address the above problems,this thesis proposes an image dehazing algorithm based on contrast learning and cycle-consistent generative adversarial network.The cyclic generative adversarial network is used to train networks with unpaired haze images and clear images to improve the application value and alleviate the problem of domain shift for image dehazing algorithms in real scenes.The contrastive guidance branch is designed to learn the potential feature distributions of images,implicitly constrain the embedding of different samples in the depth feature space,deeply mine the similar features of haze images and clear images,draw closer the similar features,retain the mutual information between the two types of images,maintain the consistency of image content,and improve the network dehazing performance.Frequency loss is used to constrain the output of the generator,reduce the frequency domain information loss,further retain the content and structure information of the image,reduce the blurring and distortion of the dehazed image,and improve the quality and clarity of the generated image.The experimental results show that the proposed algorithm is an effective image dehazing method.Information entropy and average gradient are improved by1.69%,28.67% on average,compared with DCP,CAP,Cycle GAN,and Cycle-Dehaze...