Research On Image Dehazing Algorithm Based On Generative Adversarial Network

In recent years,environmental pollution has become more and more serious,and hazy weather has become more and more frequent,which seriously affects our daily production and life.In hazy weather,optical equipment can not obtain effective scene information,resulting in serious color distortion,low image contrast and poor imaging quality,causing a great impact and interference on video surveillance,automatic driving,satellite remote sensing and other fields.Therefore,how to effectively remove haze in images without losing image details or introducing additional interference information,recover image details and texture information to the maximum extent,and improve image clarity and saturation have important research significance and application value.Image dehazing is a very challenging research topic,which has attracted extensive attention from academia and industry.However,most image dehazing methods rely heavily on the atmospheric scattering model to linearly fit haze-free images by estimating the scene depth map and atmospheric ambient light of hazy images.But in real natural environment,the cause of haze is complex,and it is difficult to be fit by a simple linear formula.At the same time,inaccurate estimation of parameters such as scene depth map and atmospheric ambient light also further influences image dehazing effect,resulting in color distortion,artifacts,as well as other phenomenon such as dehazing is not obvious.In addition,many neural network based haze removal methods apply the same weight value to different spatial pixels and different feature channels in the process of inference.However,haze is often uneven distributed across an image,and feature information extracted by different feature channels is also different.Therefore,different attentions or weight values should be applied to different spatial pixels and different feature channels across an image to improve the feature expression ability of neural network.At the same time,many neural network based dehazing methods require pairs of hazy and haze-free images in the same scene as training data,but these images are difficult to obtain in practical application,which also severely restricts the development and application of dehazing networks.It is also worth pointing out that,most of the dehazing methods based on neural network utilize the convolution operation to extract feature information.However,convolution operation shared with parameter characteristics has two disadvantages when applied,the one is that convolution operation mainly focuses on local characteristic information extraction,and not able to model feature beyond the scope of receptive field.As such,the global feature information of images cannot be perceived well.The other one is that the interaction between the convolution kernel and the image cannot be adjusted adaptively according to the image content,so it may not be the best choice to use the same convolution kernel to reproduce the image of different regions.In view of above problems,the paper researches on the problem of image dehazing.Based on the generative adversarial network and the key technologies of densely connection network,U-Net network,Transformer,residual network and attention mechanism,a series of work such as problem analysis,network design,theoretical research and simulation experiments have been carried out.The main research contents of this paper are listed as follows:· Aiming at the problem of atmospheric scattering model and neural network attention,an image dehazing method based on residual spatial and channel attention network is proposed.This method does not need to estimate any atmospheric scattering parameters,and can directly recover clear haze-free images from input hazy images.Specifically,we proposed a residual spatial and channel attention module,which can adaptively rescale weight values of different spatial pixels and channel features,based on analyzing the relationships between different spatial pixels and channel features.This makes the neural network concentrate more on valuable spatial pixels and channel features,and improves the feature expressing ability of neural network.At the same time,the contrast loss function and registration loss function are proposed to train the dehazing network,so as to better retain the details and texture information in the images and reduce the generation of artifacts.Experimental results reveal that the method achieves better dehazing effect in both public synthetic image datasets and real hazy images,and the restored images have clearer details and richer colors.· To solve the problem of paired image datasets,a haze relevant feature attention dehazing network has been designed based on the cycle generative adversarial networks.Specifically,considering various differences in physical properties between hazy and haze-free images,we first calculate the brightness,saturation,contrast,color attenuation and dark channel feature maps of hazy images,and then put them into the haze relevant feature attention networks to get the corresponding hazy attention maps.Since haze is usually unevenly distributed across an image,the hazy attention maps can provide hazy concentration information of each pixel in hazy images,so as to guide and provide reference for the subsequent dehazing network,resulting in better haze removal effect.At the same time,a color loss function is proposed to train the dehazing network to reduce color distortion of generated images.The training of the network does not depend on pairs of hazy and haze-free images.Only one randomly selected hazy image dataset and one haze-free image dataset are enough to complete the training of the network.Experimental results reveal that the method can effectively remove haze from hazy images and restore high quality haze-free images.· In order to overcome the problem that convolution operation cannot model features beyond the range of receptive field,a global and local feature fusion dehazing network is proposed.Specifically,transformer and convolution operation are preformed to extract global and local feature information of images,respectively.Then we fuse them and output,which gives full play to the advantages of transformer in modeling long-distance dependency and local perception of convolution operations,and achieves efficient feature expression.Before the final output of restored images,an enhancement module containing multi-scale patches is designed to further aggregate the global feature information and enrich the details of restored images with transformer.At the same time,a global positional encoding generator is proposed which can generate positional encodings adaptively according to the global content information of images,and then realize the two-dimension spatial location modeling of the dependency relationship between pixels.The experimental results demonstrate that the proposed dehazing network performs good dehazing performance in both synthetic and real image datasets,and the restored images are more realistic and clearer,and the detail reduction degree is high.