Design Of Defogging Algorithm Based On Deep Learning

With the rapid development of artificial intelligence technology and data science,in recent years,more and more computer vision tasks,such as traffic monitoring,Skynet security,target detection and so on,have higher requirements on the clarity of the input image.However,under the influence of suspended particles and soda in the atmosphere,the images collected by the camera are usually blurred,have low visibility,or their colors are changed due to haze,and the low-quality images seriously reduces the performance of various intelligent information processing systems.Therefore,it is very important to defog the haze image in computer vision.However,image dehazing is mathematically ambiguous in computer vision,the prior-based dehazing algorithms are strongly rely on the accuracy of the assumed image priors,and the performance of current methods based on deep learning is not stable.In this paper,based on the haze imaging model,we study the algorithms of single image dehazing.According to the limitations of the different types of dehazing algorithms,we propose a novel end-to-end single image dehazing method,which can jointly learn the transmission map,atmospheric light and dehazing all together.We put forth a 2-stage network: a physics-based backbone followed by a cGAN(Conditional Generative Adversarial Networks)refinement.The first stage learning is achieved by directly embedding the atmospheric scattering model into the network,thereby ensuring that the proposed method strictly follows the physics-driven scattering model for dehazing.For the second stage,we propose a cGAN to recover the background details failed to be retrieved by the first stage,as well as correcting artifacts introduced by that stage.Inspired by the dense network that can maximize the information flow along features from different levels,we propose a novel RDN(Residual Dense Network)to recover realistic clear images from the outputs of the first stage.To generate better results,we further modify the basic cGAN formulation by introducing the perceptual loss and smooth L1 loss.we propose a PatchGAN discriminator to decide whether the dehazed images are real or fake.In the experiment,we synthesize a hazy dataset including indoor and outdoor scenes to train and evaluate the proposed algorithm.The average PSNR(Peak Signal-to-Noise Ratio)and average SSIM(Structural Similarity)values of the algorithm on the public data set are better than the state-of-the-arts,reaching the highest 22.673 and 0.901,respectively.In the experiments on the real-world dataset,our method can also recover clear images with better details.