Image Dehazing Algorithm Based On GAN Network

Fog is a kind of atmospheric phenomenon,particles suspended in the air will strengthen the scattering and refraction of light,resulting in the sharpness and recognition of the image taken outdoors are greatly reduced,and the loss of the details of the long-distance target image,accompanied by color deviation,image artifacts and other phenomena.This has lost the authenticity of the scene to a large extent,and has a negative impact on human subjective feelings,which is not conducive to target identification and capture of details.Therefore,it has important practical significance for image sharpening processing and research.Based on the analysis of the existing image de-fogging algorithm,this paper studies and discusses the de-fogging technology based on physical modeling and deep learning,and puts forward two kinds of network models.In order to solve the problems of color distortion and residual fog after image de-fogging algorithm processing,this paper proposes an enhanced Pix2 pix HD image de-fogging method.The method consists of a multi-scale generator,a multi-scale discriminator and an enhancer.The generator is composed of global sub-generator and local sub-generator.This module extracts the feature information of different scales,so that the generated picture has both global and local authenticity.In order to enhance the transformability of the model and make it not limited to regular object when extracting features,an enhancer module is added,which is composed of deformable convolution and SKNet network.The experimental results show that the proposed method can make better use of the feature details of different scales of the image,improve the effect of image defogging,and has a certain competitiveness compared with the existing defogging methods.In order to solve the problem that existing de-fogging algorithms need to train fogged images and non-fogged images in pairs,this paper proposes a GLCGAN image de-fogging method based on residual-attention transformation,so as to solve the problem that training data need to be perfectly matched.In this method,the U-Net structure generator is improved,and the structure of the conversion network is changed into two series and then parallel.The conversion network is composed of residual blocks and mixed attention mechanism.The improved structure reduces the loss of feature information,reduces the depth of the network,increases the width of the network,and makes the network more flexible in processing features of different scale.The experimental results show that the method proposed in this paper can restore the details of the image well,and it has a great improvement compared with the existing de-fog algorithm.Figure [29] Table [2] Reference [64]...