Research On Dehazing Algorithms Based On Generative Adversarial Networks

Low visibility situation caused by dense fog trends to deeply affect vehicle and pedestrian disciplinary intelligent identification systems.Under such circumstances,small droplets and particles floating in the air scatter atmospheric light block the path of reflected light of the imaging object,reducing the reflected light received by the camera,thus the quality of imaging reduces.Therefore,image dehazing task makes a great difference to ensure the stability and effectiveness of inetlligent identification systems.This paper improves traditional haze image synthesis modeling method,in the light of fixed surveillance camera proposes a method to synthetize fog images from clear images in pairs based on dark channel for the use of deep learning network training and quality evaluation of image restoration and conduct image dehazing task based on Deblur GAN network and its improvements.The main research contents are as follows:First of all,synthetize a haze image dataset.In the light of fixed surveillance camera situation,this paper proposes a method to synthetize fog images from clear images in pairs based on dark channel,using dark channel prior algorithm to extract transmission maps of fog images,introducing a modified extinction coefficient and extracting V channel luminance matrix to model and synthesize haze images.This paper contributes a synthetic haze image dataset based on Jiangsu Tongqi expressway surveillance video frames and conducts subjective evaluation on it.Secondly,conduct image dehazing task based on Deblur GANs.In view of the traditional image processing only low resolution images,this paper conducts image dehazing task by using image deblur methods,using Deblur GAN network to dehaze fog images.In order to improve the resolution of image processing and preserve detail texture information,this paper trains Deblur GAN-v2 model on the same dataset synthetized above,reappeares Dark Channel Prior algorithm and severally test on the test dataset,and does tests on real haze datas and synthetic haze datas from Cityscape Dataset.Finally,propose Deblur GAN-Attention Algorithm to dehaze.In order to correct the colour difference in the air and preserve more detail texture information,this paper introduces Sknet module which belongs to Attention Mechanism to the FPN framework of Deblur GAN-v2 network so that it can wisely guide the network to choose the appropriate convolution kernel size and select more important feature maps when carrying out feature fusion,and do test on real haze datas and synthetic haze datas from Cityscape Dataset.