Research On Image Defogging Algorithm Based On Deep Learning

The image degradation caused by fog image imaging has a great impact on the applications of traffic monitoring,military aerial reconnaissance and target recognition based on image and video.The nature of the image to fog from degraded images to remove the disturbance of the weather,enhance image clarity and color saturation,as far as possible will be useful feature to restore the original image,the main work of this paper is the analysis of characteristics of fog figure and the fog imaging mechanism,on the basis of aiming at the existing problem of traditional to fog algorithm,This paper studies the algorithm of using deep learning form to train neural network to reconstruct original clear image.The main contents include the following points:Firstly,To gain a deeper understanding of the basic process of fog map imaging,the atmospheric scattering model under foggy conditions is studied,and the effects of incident light attenuation and atmospheric particle scattering on the image acquisition device's image acquisition are analyzed.On the basis of these theories,the classical dark primary color a priori defogging algorithm is studied,and the advantages and drawbacks of dark channel defogging are analyzed through specific experimental results.In order to make a comprehensive evaluation of the performance of the subsequent defogging algorithms,various methods of image quality assessment including subjective image quality evaluation criteria and objective image quality evaluation criteria are explored in the paper.Secondly,To address the limitations of the image enhancement class of defogging algorithms represented by the dark primary color a priori defogging algorithm in applications,such as the recovery failure of the sky region.In the context of atmospheric scattering model to build a neural network to complete the estimation of the unknown parameters in the defogging formula,the algorithm fully explores the relationship between the feature loss of foggy images in terms of brightness and saturation and the depth of the scene,combines the features of the sky region in terms of brightness,location and connectivity,uses the U-net network to complete the initial division of the image,and then uses multiple pairs of foggy clear images of the scene to complete the training of the network is then completed by using several pairs of clear images of the fog map to complete the training of the network,and a fog removal formula suitable for most scenes is fitted.The results show that this method can recover sky area better and achieve good subjective effect in improving clarity.Thirdly,In order to solve the difficulty of using neural network training that requires a large number of pairs of fog clear maps as training data acquisition,this paper uses the CycleGAN network to complete the end-to-end conversion from fog to clear maps.The cyclic symmetric structure of CycleGAN is used with the network feature of cyclic consistent loss to reduce the difficulty of training data scenes that need to be matched exactly.In order to make the network converge faster and achieve the desired effect with certain robustness,this paper improves on the CycleGAN by transforming the form of residual network with more diverse loss functions with different weight values to modify the network training effect.The results show that the algorithm of this paper can recover more image information in the depth of the scene than the traditional algorithm,and the color fidelity of the sky area is high.