Research On Image Dehazing Based On Dark Channel Prior And Deep Neural Network

In recent years,as the problem of environmental pollution has become increasingly severe,the frequency of smoggy weather has also increased day by day,which has caused great obstacles to people’s daily travel.In foggy weather conditions,due to the inability of optical instruments to obtain effective environmental information,the color of the image is distorted,the contrast is reduced,and the image quality is poor.These problems have brought great difficulties to fields such as video surveillance,unmanned driving and navigation.Therefore,how to effectively eliminate the fog without losing image texture details or other interference information,and restore the original features of the image as much as possible,so as to improve the clarity and saturation of the image,has very important research value.At present,the research on dehazing based on deep learning is one of the important directions on improving the quality of image dehazing.Considering the challenges of detail loss and color distortion after image dehazing,this paper conducts the following researches on the image dehazing problem based on deep neural network:(1)A multi-scale feature fusion dehazing network combined with a continuous memory mechanism is proposed.First,the network model obtains multi-scale feature maps through downsampling operations,and further employs skip connections between corresponding network feature layers to connect the feature maps between the encoder and decoder to achieve good feature fusion.Then,a continuous memory residual block is introduced to enhance information flow and enable feature reuse.In addition,to utilize the texture representation information and complete adaptive dehazing according to the haze density,this method introduces a pixel attention module on the skip connection to combine the residual dense module of the corresponding decoding layer,thereby improving the feature expression ability of the network.Experimental results show that the proposed model can achieve better dehazing performance compared to other state-of-the-art methods based on deep neural networks.(2)A feature fusion dual-branch dehazing network based on dark channel prior and Transformer is proposed.To address the feature inconsistency between Transformer and CNN,the network modulates CNN features by learning a modulation matrix conditioned on Transformer features,which naturally inherits the global modeling ability of Transformer and the local representation ability of CNN.At the same time,through a three-dimensional position embedding module based on the scene transmission map perception of the dark channel prior,the prior related to the haze density is introduced into the Transformer.This module provides the relative position,and prompts the haze density in different spatial regions.Specifically,the network uses the Transformer branch and the contextual convolution operation branch to extract image global and local feature information,respectively.After the two are fused,the final restored image is output,which fully utilizes the advantages of the Transformer’s global modeling ability and the local perception characteristics of the convolution operation,and realizes the efficient expression of features.Experimental results show the haze-free image reconstructed by the proposed algorithm has high clarity and saturation and the dehazing effect is excellent.