Research On Image Dehazing Algorithm Based On Deep Learnin

In recent years,deep convolutional neural networks(DCNN)have been widely concerned and achieved remarkable results in image dehazing.In the high-level version tasks,such as object detection and image classification,such tasks need clean image input,so as to ensure the accuracy of model prediction.However,in harsh weather conditions such as smog,photos taken by cameras often show low contrast and poor visibility.Therefore,the purpose of image dehazing task is to greatly restore the ideal clear image and improve the image quality.However,existing dehazing algorithms still suffer from color distortion,loss of edge details,and low image recovery metrics.To address these problems,two novel dehazing algorithms are further proposed in this paper,and the main works are as follows:(1)A multi-scale feature dehazing network(HF-Net)based on high-frequency feature fusion is designed and applied to single image dehazing.HF-Net uses multi-scale feature network as the backbone,which can share effective feature across different scales.In addition,the Laplacian operator is used to obtain the high-frequency information of the image.The highfrequency information can representatively represent the texture details of the image,and then is fed into the frequency branch network as an additional prior to extract the high-frequency information features.A specially designed frequency attention module(FAM)was embedded,which can autonomously learn the weights map of high-frequency features to enhance the model recovery ability.Finally,with the help of the refine module(RB),multi-scale features and high-frequency features are integrated to generate clear images.(2)A color-guided and edge-guided dehazing network(CEG-Net)is designed to alleviate the problems of color distortion and edge detail blurring in many existing dehazing methods.CEG-Net consists of a cross-aggregation dual U-Net(CADU),a color feature extractor(CFE),a prior attention module(PAM),and a prior fusion group(PFG).Specifically,CADU can take advantage of the complementarity between edge and texture features in a cross-aggregation manner.CFE focuses on hue and saturation channel characteristics to address color distortion under color correction loss constraints.Finally,PAM and PFG are used to handle different prior factors(i.e.,edge,texture,color),and the attention mechanism and fusion strategy were used to guide the feature learning process,respectively.