Research On Image Dehazing Algorithm Based On Semi-supervised Learnin

With the development of computer vision technology,related achievements have been widely used in many fields such as medical treatment,transportation,and national defense.However,under the influence of severe weather such as haze,the images obtained by imaging equipment are often of poor quality,which in turn affects th e performance of related technology applications.Therefore,image dehazing has become a crucial problem and is now a hot area of research.However,the current dehazing datasets are all synthetic hazy images,which are quite different from real hazy images,which limits the performance of deep learning dehazing algorithms to a certain extent.In order to solve the problem that real hazy images cannot be applied to network training,this paper proposes two different image dehazing algorithms based on semi-supervised learning:(1)A semi-supervised image dehazing algorithm based on generative adversarial networks.Using Generative Adversarial Networks as an unsupervised branch,real hazy images are introduced into model training,and the networks are trained alternately with synthetic datasets.Smooth dilated convolutions are introduced in the network to improve the network receptive field and eliminate grid artifacts.The discriminator of the generative adversarial network adopts the Markov discriminator to improve the recovery ability of the network to details.(2)A semi-supervised image dehazing algorithm based on self-supervised learning.Considering the difficulty of generating adversarial network training,the unsupervised branch is changed to a self-supervised learning algorithm.In the unsupervised branch,the haze residual K of the hazy image is estimated by AOD-net,and the dehazed image is restored to a hazy image by the improved atmospheric scattering model in AOD-net to achieve the purpose of self-supervised learning.In order to verify the effectiveness of the two algorithms,this paper compares the SOTS test set in the RESIDE dataset with five classical algorithms,and selects some images for comparative analysis.And in order to verify the effectiveness of semisupervised learning,ablation experiments are carried out.The experimental results show that the two algorithms proposed in this paper are superior to the comparison algorithms in both PSNR and SSIM indicators,and have improved in terms of dehazing degree and texture clarity.generalization ability.