The Research Of Image Dehazing Algorithm Based On Deep Learning

In recent years,with the rapid development of artificial intelligence and big data,more and more intelligent information processing systems,such as traffic monitoring systems and unmanned systems,have imposed higher requirements on the clarity of input image.However,due to the bad weather conditions such as haze and fog,the quality of obtained image is generally low,and there are often phenomena of low contrast,hue shift,and low visibility of information.These degraded pictures not only affect the subjective feelings of the human eyes,but also seriously affect the performance of various types of intelligent visual information processing systems.Therefore,it is very important practical application value to clear the fog image.However,the image dehazing is a very challenging ill-posed problem,which has drawn more and more scholar's attention.On the basis of fully understanding the atmospheric scattering model,this article studies on single image dehazing algorithm.According to the limitations of different types of dehazing algorithms,two different algorithms based on deep convolutional neural network are proposed and achieve superior dehazing performance than the current state-of-the-art methods.The contributions of this article are mainly as followings:(1)Research on image dehazing algorithm based on deep full convolutional regression network.In view of the fact that the performance of existing prior-based image dehazing methods is limited by the effectiveness of prior knowledge,and the recovered haze-free images by the existing CNN-based dehazing algorithms sometimes are still unsatisfactory.This article proposes an image dehazing algorithm based on a deep full convolutional regression network.The network learns intrinsic mapping between the input hazy images and their corresponding transmission,which predicts accurate estimation of the transmission of the scene.Moreover,in order to optimize the network training process,a training data synthesis method based on unpaired images is proposed,which avoids the strict dependence of training data on scene depth pairing images.Through the synthetic data that is closer to the real fog map,the network has obtained a good training performance.Qualitatively and quantitatively experimental results on the synthetic and real-world hazy images demonstrate that the proposed method effectively predicts more accurate transmission map and removes haze from hazed images,The recovered image overcomes the problem that the traditional dehazing algorithm is unsatisfactory for sky area recovery and is natural and visual pleasing.(2)Research on the dehazing algorithm for joint estimation of transmittance and atmospheric light.For the existing CNN-based algorithm,only the transmission rate estimation through the network is achieved,and no more effective atmospheric light value calculation method is proposed in combination with the network.This article also proposes a new network of joint estimation of transmittance and atmospheric light values.The algorithm uses CNN to estimate the parameters of the atmospheric scattering model and avoids the problem of the inaccurate estimation of the atmospheric light value by the traditional method.Based on the use of low-level shared feature units,the network structure extends the transmittance estimation and atmospheric light value estimation branches respectively,and through the joint estimation method,the integrity of network training is ensured.The experimental results on the synthetic and real-world hazy images demonstrate that the proposed joint estimation method outperforms other CNN-based algorithms.