Research On Image Dehazing Method Based On Deep Learning

With the advent of the information technology,a variety of imaging equipment has been spread in every corner,the quality of images is an important prerequisite for the success of information technology systems.The existence of haze and smog makes the images obtained by imaging devices hazy and unclear,which directly affects the further use of images by information technology systems,so there is a practical need to recover high quality clear images from hazy images,and the study of dehazing algorithms has realistic use value.This paper starts from the development history and research status of dehazing algorithms,studies the theoretical knowledge and technical basis commonly used in image dehazing algorithms,research and compares six classical dehazing algorithms,finds some advantages and disadvantages of them,and summarizes four experiences by comparing and analyzing their dehazing effects to provide reference for the subsequent research of dehazing algorithms.With the experience summarized from classical dehazing algorithms,this paper propose a dehazging network named C-DehazeNet based on the Encoder-Decoder structure,combining multi-scale fusion techniques,feature attention mechanism,deformable convolution and contrastive regularization.The translation network CUT is selected to extend the dataset for alleviating the domain shift problem.An image dehazing system is designed and implemented,integrate the dehazing methods studied in this paper into the system.The main research is as follows:1.Studies the evolution of the main algorithms in the development of dehazing methods,select six of the classical algorithms for research,analyze their advantages and disadvantages by comparing each network itself and comparing different networks with each other,and summarize four experiences that can be referred to when designing dehazing networks.2.A network named C-DehazeNet is proposed.The network is trained using OTS data sets,the dehazing results of C-DehazeNet and the classical dehazing algorithm is compared from a subjective point of view,the effectiveness of the C-DehazeNet is verified using objective evaluation indexes,and the effectiveness of each module of the C-DehazeNet is demonstrated by ablation experiments.3.Most dehazing algorithms are trained using synthetic datasets,and the models are not effective for real hazy images,those phenomenon of training models on synthetic domain to process real domain images poorly is called domain shift problem,and to alleviate this problem,this paper uses image translation network CUT to translate the hazy image datasets to each other to bridge the domain shift gap.4.A computer image dehazing system is designed and implemented.The system integrates six classical dehazing methods and the C-DehazeNet proposed in this paper,and user can select the dehazing method and the corresponding model for dehazing according to the need.In this paper,using the Outdoor Training Set(OTS)in RESIDE and the translated dataset train the proposed C-DehazeNet.Then,use the real hazy image and the Synthetic Objective Testing Set(SOTS)to evaluate the dehazing effect of the C-DehazeNet;use Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity(SSIM)as objective evaluation indexes to evaluate the C-DehazeNet.Compare the color reproduction,contrast,brightness and local detail recovery of the dehazed images to evaluate the degazing results from a subjective perspective.