End-to-end Nighttime Image Dehazing Research

With the acceleration of science and technology,the development of computer vision applications,such as driverless vehicles,road monitoring,remote sensing satellites,etc.,has put forward higher requirements for images quality.However,the southeast coastal areas due to high air humidity and many dusts,resulting in many haze weathers.The existence of haze makes the clarity of the image has a serious decline,which hinders the processing and judgment of intelligent information acquisition equipment.The nighttime is greatly affected by lights,and the pixel values of the whole scene are lower than daytime.Combined with the influence of haze,it becomes more difficult to remove haze at night.In view of the current results of nighttime dehazing,there are problems such as low degree of visualization,color distortion and halo artifacts,as well as a few fast image dehazing methods.Based on the knowledge of prior,physical model and convolutional neural network,this paper proposes a new endto-end nighttime image dehazing method.In order to achieve the purpose of image clearness,this dehazing method is to improve the images brightness and remove the haze from images,respectively.(1)Since it has been discussed in Maximum Reflectance Prior(MRP)that the ambient illumination at night is almost irrelevant with color.This paper aim to eliminate the influence of color and improve the visualization of the result of haze-free images,Retinex theory,Maximum Reflectance Prior(MRP)theory,and Guided Filter(GF)algorithm are combined to implement color correction of hazy images.Through this processing of color correction,the brightness of the dehazing results enhance a lot.(2)In view of the current night haze removal methods exist problems of overestimate or underestimate the transmission maps,an autoencoder with skip connections is proposed to estimate the transmission maps,which can improve the accuracy of the transmission maps.In order to realize the estimation of the ambient illumination,a CFNet(Context Fusion Network)is proposed.This network according to Guided Filter which keeps the edge of low frequency images and Autoencoder which has Smoothed Dilated Convolution and Gated fusion to learn the residual filtering results between the filtering results of hazy images and the filtering results of haze-free images.In this method,the filtering result of haze image with light source and the residual image without light source are combined to realize the accurate estimation of ambient illumination.For the input hazy images,transmission maps and ambient illumination maps,the atmospheric imaging model in the end-to-end nighttime image dehazing method can be used to restore clean haze-free images.Because the atmospheric scattering model analysis the cause of haze from the physical aspect,the result of haze removal obtained in this way conforms to the physical acknowledge.(3)At present,there are few methods to synthesize the nighttime haze datasets,and these methods do not take the changing light source location into account.In this paper,we propose a synthetic method of darkening different image brightness,randomly setting light source position and setting ambient light with different intensity.According to this synthesis method,the hazy images at night are more suitable for the actual situation at night scene,which also makes the synthesized data more robust.(4)In the experimental part,in order to highlight the effectiveness of the method proposed in this paper,a large numbers comparative experiments are carried out between the proposed method and the state-of-the-art methods.The results show that our method can effectively remove haze?suppress the hole artifacts?obtain more image details and can quickly remove haze from input images.In addition,the average Peak Signal to Noise Ratio(PSNR)is 21.2512 and the average Structural Similarity(SSIM)is 0.5943.Both them are higher than those state-of-the-art methods.