Research On Image Dehazing Method Based On Attention Mechanism And Semi-supervised Learnin

Fog and haze float in the air in the form of fine particles.It scatters the natural light in the air on the one hand and absorbs the reflected light generated by the sunlight shining on the objects on the other hand,resulting in contrast degradation and color distortion in foggy images,which seriously affects the application effect of using image information as input devices.Therefore,in order to maintain the normal operation of the relevant vision system under foggy conditions,the study of image defogging technology is of great importance.The current image defogging methods are mainly divided into traditional defogging methods and deep learning-based methods.Traditional defogging methods include image enhancement and physical model-based defogging methods.Deep learning methods have better defogging effects and are more robust than traditional methods.Therefore,this paper adopts deep learning methods for image defogging techniques,and the main contents of this work are as follows:(1)An end-to-end image defogging method based on the attention mechanism is proposed to address the problems of current image defogging methods that directly rely on atmospheric scattering models,lack of intermediate quantity estimation and network under-constraint.The method first embeds the channel attention mechanism into the inception network,and the shallow features are extracted by the fused network;then the deep image information is learned by the multi-scale convolution module and the residual dense connection block,and the deep and shallow features are fused by jump connection;finally,the single convolution layer is regressed to the pixel scale factor matrix P(x),and the fog-free image is restored based on the improved atmospheric scattering model.The fog-free image is restored based on the improved atmospheric scattering model.In addition,the network model is additionally designed with a fidelity loss function as a constraint on top of the mean squared error(MSE)to improve the fog removal performance.The analysis of the experimental results of RESIDE foggy day dataset shows that the proposed method achieves high scores in the image quality evaluation indexes such as peak signal-to-noise ratio(PSNR),structural similarity(SSIM),learning perceptual image block similarity(LPIPS)and CIEDE2000,indicating that the recovered images are of good quality with high color fidelity and complete detail information retention.(2)A semi-supervised adversarial learning method for image defogging is proposed to address the problem that the current stage of deep learning-based defogging methods are usually only able to train on artificially synthesized foggy-sky datasets,resulting in weak model generalization ability.The method consists of a fully supervised branch and an unsupervised branch.The generative adversarial network is used as the unsupervised branch and an unsupervised loss function(dark channel loss,full variance loss)is introduced to enable the network to be trained on real foggy images without label pairs,and then during the training process,a discriminator is used to determine whether the recovered fog-free image is a high-quality clear image.To make the model more stable,synthetic fog images are added to the fully supervised branch,and end-to-end defogging is achieved by alternating training with fully supervised and unsupervised branches.The experimental results show that the proposed semi-supervised adversarial learning defogging method is not only effective on simulated fog maps,but also has excellent defogging effect on real foggy images,showing stronger model generalization ability.