Research On Adaptive Dehazing Algorithm For Single Image

Fog environment will seriously lead to the decline of outdoor image contrast,loss of color fidelity and edge information,which are not conducive to the application of computer vision(such as traffic monitoring,automatic driving and object recognition).With the rise of deep learning and more and more scholars studying dehazing algorithms,many image processing related technologies and important theories can effectively remove haze and restore high-quality images.However,in the real scene,there are not only globally distributed haze,such as fog,but also locally distributed haze,such as evaporation haze after rain.Therefore,it is of great practical significance to remove the local haze and global haze and improve the image quality.Existing dehazing algorithms can only achieve global haze removal.Due to the local haze and the global haze in the image have different characteristics.For the internal part of the local haze,it is subject to uniform distribution,while the global haze is subject to sparse distribution.Therefore,it is necessary to study local haze or global haze removal algorithm in images respectively for the problems of local haze or global haze removal.To solve the above problems,an end-to-end adaptive dehazing algorithm for single image is proposed,which automatically judges the input image type(local fog,global fog)according to the proposed region segmentation discriminative model,and adaptively selects the corresponding algorithm to achieve local haze or global haze removal of the image based on the judgement result.The main research work is as follows:(1)To determine the type of haze in the input image,a mathematical model that can distinguish the haze type in the input image is established.Firstly,according to the strong edge location ability of Laplace operator,a discrimination method based on Laplace operator is proposed.The method can effectively judge the local haze or global haze in a single image when there are many dark areas.However,this method does not fully consider the bright areas of the image.Considering the problems of the discriminant method based on Laplace operator,then a discrimination model of dark channel and color attenuation theory is proposed,which can adaptively judge the type of input image according to the distribution of haze in the image.(2)For the problem of local haze removal,an attentive generative adversarial network dehazing algorithm is proposed.The global haze removal algorithm of a single image to process the local haze image will lead to the problem of blurring the edge and degrading the overall image quality.Structural loss,total variation loss and perceptual loss are add into the algorithm for optimizing model performance in traditional generative adversarial network losses.The attention mechanism of this algorithm can focus on the area of local haze distribution and process the haze in this area to realize image dehazing.(3)For the problem of the deep learning method with huge model parameters reduces efficiency for global haze removal.Therefore,a lightweight adaptive multi-scale Retinex dehazing algorithm is proposed.The algorithm can use lightweight network to realize image dehazing,and can make up for the problem of dark image color and fog residue in image processing.Experimental results show that the algorithm can effectively remove local haze and global haze in the image.