Research On Image Defogging Algorithm Based On Deep Learning

In recent years,with the aggravation of heavy fog,the images acquired by surveillance cameras become fuzzy,and it is difficult to accurately track and recognize the target images.With the increasing demand of automatic driving,abnormal condition monitoring and atmospheric environment detection,the analysis and processing of fog image has a broader market application prospect.The infrared camera has strong penetrability,but it is expensive,difficult to maintain and has limited economic benefits.Therefore,the research of image defogging algorithm has very important theoretical research significance and practical application value.Firstly,we analyze the image defogging algorithm based on dark channel prior.Based on the prior dark channel defogging algorithm,through the study of the physical model of fog imaging in the atmosphere,the dark channel image is calculated and the atmospheric light intensity in the image is judged,and the atmospheric transmittance of the output image is obtained.However,the estimation of atmospheric transmittance is not accurate,and there is obvious halo phenomenon in the defogging image.With the continuous development of deep learning algorithm,convolutional neural networks(CNN)has an advantage in estimating the atmospheric transmittance of images.The generative adversarial networks(GAN)can get rid of the constraints of atmospheric physical model and avoid the problem of parameter error accumulation in atmospheric physical model.Therefore,this paper focuses on image defogging algorithm based on convolution neural network and generation countermeasure network.The following is the main work and innovations of this paper:1.An image defogging algorithm based on multi-scale convolutional neural network(MSCNN)is proposed in this paper.The multi-scale structure is introduced into the convolutional neural network,and two convolutional neural networks are used to output the atmospheric transmittance image.Two convolution neural networks extract the image information of coarse scale and fine scale respectively to obtain the fine atmospheric transmittance image.The deepening of the network can mine the deep information of the image,make the atmospheric transmittance image more accurate and eliminate the halo phenomenon.In addition,in the estimation of atmospheric light intensity,the super pixel segmentation algorithm is used to extract the sky area,which limits the extraction range of atmospheric light intensity to the sky area,so as to prevent other targets and point light sources in the image from affecting the estimation of atmospheric light intensity.The experimental results show that the proposed method is effective.2.Aiming at the problem of error accumulation in the process of modeling atmospheric physical model,this paper proposes an image defogging algorithm based on Wassertein GAN.The input fogged image first enters the generation network for defogging,and then the defogged image is input into the discrimination network.The smaller the earth mover distance is,the closer the defogged image is to the real image.In this paper,we use the idea of residual network for reference,and introduce jump connection into the structure of full convolution neural network to solve the problem of gradient dispersion caused by the increase of neural network depth.In order to ensure the precision and speed of the network defogging process,batch normalization is introduced to help the neural network to accelerate convergence.At the same time,the perceptual loss function is introduced to help improve the detail features of the defogging image.The experimental results show that the image defogging effect of this method is better and the performance is more stable in the case of complex background and high fog concentration.