Image Dehazing Based On Generative Adversarial Learning

Image haze removal technology has important applications in outdoor monitoring system and UAV aerial photograph system.The image acquired by outdoor imaging equipment in hazy weather is not clear and has low constract.The degradation of image seriously affects the subsequent computer vision tasks.Therefore,it is worthy to study how to remove the haze in images.Traditional image dehazing methods usually directly enhance the hazy image,or recover the clear image based on physical model and certain assumptions prior.The performance of these methods is limited by hand-designed features and are prone to image distortion,halos,and low brightness.The existing image dehazing methods based on deep learning directly learn the mapping between the hazy image and transmission or clear image,and get better feature than traditional dehazing methods.However,these deep learning based methods essentially use the convolutional network to estimate the parameters in physical model,and then recover clear image based on physical model.And only the mean square error(MSE)loss is constrained in the optimization.So the performance of these methods is not stable enough.This paper focuses on the research and improvement of image dehazing methods based on convolutional neural networks.The main work and contributions are as follows:1.An image dehazing method based on conditional generative adversarial network is proposed.Firstly we design an end-to-end dehazing network based on a convolutional neural network which directly learns the mapping between hazy and clear images and avoids the extra estimation of model parameters.Secondly not only the MSE loss but also the perceptual loss is constrained in the optimization.Then based on this,we optimize the dehazing network through conditional generative adversarial learning.Concretely,a discriminator is built to classify the dehazed images and the clear label images.The dehazed image can be more similar with the label image through minimizing the adversarial loss.The experimental results show that the performance of our methods has higher contrast and more detailed information,and the overall visual quality of the image is improved.2.An image dehazing method that combines weakly supervised adversarial learning is proposed.The existing image dehazing methods based on deep learning(including the method we proposed in contribution 1)supervised learn the map between hazy image and transmission or clear image based on a large number of synthetic hazy images.But the real hazy images are more complex than synthetic hazy images in color,contrast and luminance,so these supervised learning methods always have bad performance on real hazy image.To solve this problem,we add the real hazy images to the optimization of dehazing network based on weakly supervised adversarial learning.Concretely,another discriminator is built to classify the real dehazed images and the high quality images,so that the dehazed images can learn the feature of high quality images;and a dual network was employed to ensure constancy.The experimental results show that the dehazed result of real hazy image can be closer to the natural clear image in terms of contrast,brightness and dynamic range.