Research On Image Dehazing Algorithm Based On Convolutional Neural Network

Smog is caused by tiny particles suspended in the air.With the rapid development of data science and artificial intelligence,tasks based on computer vision processing systems such as security monitoring,intelligent driving,license plate recognition,military surveys,forestry early warning,etc.have higher and higher requirements for image clarity.Smog the existence of this makes the images acquired by the image acquisition system appear reduced in visibility,blurry whitening,and decreased color saturation,which seriously hinders the effective progress of subsequent tasks based on the computer vision processing system.Therefore,it is very important to effectively defog the foggy image so that the computer vision system can work normally in the presence of fog and haze.The existing image dehazing algorithms are mainly divided into dehazing algorithms based on physical models and end-to-end dehazing algorithms.The dehazing algorithm based on the physical model calculates the transmittance and atmospheric light value of the foggy image according to the atmospheric scattering model,and then obtains the fog-free image,and the calculation of the transmittance and atmospheric light value is very easy to cause the calculation of the overall model parameter.Imbalance,resulting in incomplete dehazing,color distortion and other poor dehazing effects.Based on the end-to-end dehazing algorithm,the foggy image is used as input,and the convolutional neural network is used to learn the mapping relationship between the foggy image and the fogless image,and then the fogless image is directly obtained without calculating the transmittance and atmospheric light value.In order to avoid problems such as incomplete dehazing and color distortion,the end-to-end processing method makes the generalization ability of the network stronger and the dehazing efficiency higher.This paper adopts an end-to-end dehazing algorithm.The main work is as follows:1.Design a dehazing network that combines attention mechanism and multi-scale features(AT-Net)Aiming at the problems of incomplete dehazing and color distortion based on the end-to-end dehazing network,this paper designs a dehazing network AT-Net that integrates attention mechanism and multi-scale features.First,a feature attention module is designed according to the attention mechanism,and different weights are generated according to different features to improve the feature expression ability of the network;then,a multi-scale convolutional layer,local residual learning structure and The feature is the basic module composed of the attention module;finally,the basic module is combined with the global residual learning structure,and the dehazing network AT-Net is designed to achieve end-to-end dehazing.The experimental results on the RESIDE data set show that the dehazing network effectively solves the problems of incomplete dehazing and color distortion by fusing the attention mechanism with multi-scale features,and achieves a good dehazing effect.2.Designed a densely dilated convolutional neural dehazing network(DC-Net)Aiming at the problem that the end-to-end dehazing network is easy to lose detailed information and high computational cost,this paper designs a densely expanded convolutional neural dehazing network DC-Net.First,a dense expansion module is designed by using dilated convolution with different expansion rates;then,a three-row and six-column network composed of the above-mentioned dense expansion module and feature attention module is designed;then,in order to improve the network performance The fog capability divides the network into a pre-processing module,a backbone module,and a post-processing module;finally,a loss function combining smooth L1 loss and perceptual loss is used.The experimental results on the RESIDE and D-HAZY mixed data set show that the DC-Net dehazing network uses expanded convolution instead of ordinary convolution,which effectively solves the problems of loss of detailed information and high computational cost,and verifies that DC-Net dehazing The effectiveness of the network.This paper takes foggy images as the research object.Based on the end-to-end dehazing algorithm,the image dehazing network is designed and built based on the convolutional neural network,which achieves a good dehazing effect and effectively solves the problem of end-to-end dehazing.The dehazing algorithm has unsatisfactory dehazing effect,easy to lose detailed information,and high computational cost.