Research On Single Image Haze Removal Based On Deep Learning

Single image dehazing aims to recover the uncorrupted content from a hazy image and restore a clean image.Traditional defogging algorithms have problems such as incomplete defogging and color distortion.Image dehazing continues to be one of the most challenging inverse problems,which has drawn much scholar’s attention.In recent year,deep learning can handle many image-related visual tasks and has shown relatively advanced performance.In order to improve the defogging effect of a single image,this article will use deep learning to deal with this problem.The specific research contents are as follows:Firstly,a single image dehazing algorithm based on YCb Cr fusion residual dense network is proposed.This paper uses convolutional neural network to dehaze the end-to-end image of the brightness channel of YCb Cr color space,which can simply extract the foggy area of the image and enhance the visual contrast.In addition,the residual dense module and multi-scale enhancement module are added to the network structure,which can fully extract multi-level feature information.The experimental results show that the proposed algorithm achieves better defogging effect.Secondly,in order to further improve the performance of the single image defogging algorithm and eliminate the influence of halo artifacts on visual effects,this paper proposes a defogging algorithm based on multi-scale residual dense network.The algorithm first estimates the rough transmission map of the foggy image based on the prior knowledge of the dark channel,and inputs it into the multi-scale defogging network constructed by the residual dense attention module.Then through the atmospheric scattering model,the fine transmission map obtained by the network training is deduced into a clear fog-free image.Experimental results show that this method improves the effect of image dehazing and makes the restored image have better visual quality.Finally,in order to improve the accuracy of estimating the haze and reduce the spatial redundancy in deep learning,a single image dehazing network based on multi-scale octave convolution is proposed.Based on the idea of image decomposition,this paper designs a multi-scale octave convolution structure and integrates it into the encoder-decoder defogging network.Experiments show that this method has the best defogging effect,it can better retain the structure and texture in the image,and reduce the redundant information in space.