Research On Image Dehazing Algorithm Based On Deep Learning

With the development of industry,transportation and other fields in China,haze weather frequently occurs in North China and Northeast China,especially during autumn and winter festivals.Haze weather not only causes great troubles to people’s lives,but also brings many challenges to tasks based on computer vision processing systems.Due to the presence of haze,the collected image data shows problems such as white blur,poor visibility,and decreased saturation,which hinders the progress of subsequent tasks.So,it is of great significance to remove the foggy parts from foggy images and restore non foggy images,so that computer systems can work in the presence of haze.In recent years,research on haze removal has grown rapidly,and many research institutions have launched unique haze removal algorithms.Among them,the algorithm based on physical models calculates the image under the influence of haze based on the atmospheric scattering model,first calculating the atmospheric light value and transmittance,and then using the scattering model to calculate the fog free image.The algorithm based on deep learning can be divided into two directions:firstly,based on the atmospheric model,various parameters are calculated through neural networks to obtain fog free images;Another approach is to input foggy images and use neural networks to learn the mapping relationship between foggy and non foggy images,ultimately achieving end-to-end defogging effect.However,existing defogging algorithms still have problems such as incomplete defogging,distortion and blurring of the image after defogging,and excessive computational costs.In order to solve the above problems,this article designs two defogging algorithms based on convolutional neural networks,and the main research content is as follows:Designed a DA-Net defogging network: Currently,most end-to-end defogging networks have problems such as incomplete defogging images and susceptibility to color distortion.To address these issues,a DA-Net defogging network based on attention and multi-scale convolution was designed through spatial attention,channel attention,and multi-scale features.Firstly,a feature attention module was designed,which is derived from the attention mechanism and generates different weights based on different features.This module enhances the network’s feature expression ability;Then,a basic module was designed,consisting of multi-scale convolutional layers,local residual structures,and feature attention modules;In order to further improve the defogging effect,we combined this basic module with a global residual learning structure to design a DA-Net defogging network,achieving end-to-end defogging.In DA-Net networks,multi-scale convolutional layers can better control the learning granularity,local residual structures can fully utilize local information of images,feature attention modules effectively improve the learning effect of global information of images,and global residual learning structures can better promote the flow of gradient and information transmission,making the learning effect of the entire network more outstanding.In order to solve the problems of loss of detailed information and high computational cost faced by end-to-end dehazing networks,a densely expanded convolutional neural dehazing network,DDC-Net,is designed.In order to improve the dehazing effect of the network,we divide the network into pre-processing module,backbone module and post-processing module.The most basic module in DDC-Net is the dense expansion module,which uses the expansion convolution design with different expansion rates to effectively preserve the details of the image.Next,we combined the densely expanded module with the feature attention module in DA-Net to form a three-row,six-column network.In DDC-Net,we optimize the network by combining the loss function of smooth L1 loss and perceived loss,thereby improving the robustness and accuracy of the network.By replacing ordinary convolution with extended convolution,DDC-Net can effectively solve the problems of loss of detailed information and high computational cost.Through subjective and objective evaluation,the effectiveness of DDC-Net dehazing network is verified.