Image Dehazing Research Based On Transmittance Correction And Deep Learning

In the weather with haze,there are losts of suspended particles in the air,the reflected light of the scene will pass through these particles and scatter before reaching the imaging equipment,resulting in the collected images which cannot get effective distant vision information getting severely degraded.As a result,the application value of images is greatly reduced,which has a great impact on industrial production,security monitoring,military investigation and daily life.However,the existing dehazing methods have various problems such as distortion and non-uniformity of dehazing results.Based on the analysis of hazy image imaging model and deep learning model,this paper studies the image dehazing method from two directions: dark channel prior and deep learning.Firstly,in terms of the dark channel prior dehazing method,based on the imaging mechanism of the atmospheric scattering model and the research of the dark channel prior dehazing method,an improved method based on the channel transmittance non-consistency correction is proposed in this paper to solve the problem that the results of the dark channel priori dehazing method are prone to color distortion.The difference of RGB incident light frequency makes the atmospheric scattering coefficient of each channel different,which leads to the different transmittance value of each channel.In this paper,the optimum atmospheric scattering coefficient ratio is obtained through the dehazing experiment on the O-haze data set,and the relationship between the transmittance of each channel is then derived.Then the corresponding transmittance is used on each channel to restore the image.Experimental show that the image color of the result is more real and natural and this algorithm can effectively improve the color distortion problem of the priori algorithm of the original dark channel.Secondly,in the aspect of image dehazing based on deep learning,this paper implements an end-to-end image dehazing network based on Dense Net and Auto-encoder.Feature extraction and feature reconstruction are carried out by encoder and decoder,and multi-scale features are fused by a multi-level pyramid pooling module to further improve the representation ability of the model.The concat between the corresponding feature maps of the encoder and decoder and the dense connection structure inside the dense block enhance the propagation of features and backpropagation of the gradient in the whole network,alleviating the problem of gradient disappearance of deep network during training.The problem of gradient disappearance of deep network during training is further alleviated and the network can be successfully trained.Experiments with a large number of data show that the proposed algorithm can achieve competitive results than the current first-class algorithm.