Research On Context Aggregation And Domain Adaptation Based Image Dehazing

Most dehazing methods exist some problems.For example,uneven dehazing,highly dependence on labeled synthetic datasets and poor generalization.In view of the abovementioned problems,in this thesis,we mainly study the following two aspects:1.An image dehazing method based on a recurrent context aggregation network is proposed.Most existing image dehazing methods exist the problem of uneven dehazing.In order to alleviate this problem,we design a context aggregation block to combine global and local features.This block uses global features to provide the global visual perception of an image,and then it uses local features to restore the local details of an image to refine the global visual perception of an image,which can effectively alleviate the problem of uneven dehazing.Furthermore,we propose a deep recurrent mechanism,so that the network does not introduce new parameters while increasing its depth.By doing this,we can alleviate the problem of overfitting and improving the generalization of the model.The experimental results on the benchmark dataset show that the proposed method can effectively dehaze,and the images after dehazing are clear and natural.2.A semi-supervised image dehazing method based on domain adaptation is proposed.In order to alleviate the problem of domain shift between synthetic datasets and real datasets,and improve the generalization of dehazing methods,we propose a domain adaptive semi-supervised image dehazing method.The method first separates the content features and haze features of a real hazy image and a synthetic hazy image,respectively.Then the separated content features are fused with the haze features of different domains.Finally the fused features are sent to a dehazing network for dehazing.The proposed method alleviates the problem of domain shift between real datasets and synthetic datasets,and improves the generalization of the dehazing model.The experimental results on benchmark datasets show that the proposed method is superior to the state-of-the-art semi-supervised dehazing methods.And it can stably and effectively dehaze images on different datasets.