Single Image Dehazing Algorithm Based On Domain Transformation For Image Matching

Haze and fog are common weather phenomena.The brightness and contrast of images captured in hazy environments are always reduced,producing the degraded images,which lose a lot of information and seriously affect the performance of subsequent computer vision tasks.As a fundamental and critical task in computer vision applications,image matching relies on image feature information.When the input image is a foggy or haze image,key points cannot be detected accurately because of the serious loss of image feature information,which results in the degradation of image matching performance.Consequently,it is very important for image matching to restore feature information of hazy images.However,most of existing networks only focus on the information recovery in the pixel domain and ignore the feature domain information.Besides,when networks are directly transplanted to the real scene,the performance of networks trained on the synthetic dataset will decline.Therefore,it is extremely necessary to design a single image dehazing algorithm to recover more useful feature information for dehazed images,which is of great benefit to hazy image matching.To solve the problem,in this paper we propose a multi-task learning image dehazing algorithm to effectively recover salient features of dehazed images for image matching.The contributions of our work are described as follows:(1)We first propose a single image dehazing network aiming at retaining feature information.In this network,not clear image but its Gaussian blurred versions at multiple scales are learned.These Gaussian images are used to construct the scale space for feature point detection,and we can recover more key points in the feature domain.(2)By using domain transformation network,the mapping relationship between real domain and synthetic domain is studied to bridge the gap and enhance the dataset.(3)We design a new loss function based on the Alternative Interest Point(ALP)detector,called Scale Space Response(SSR)loss.Using this new loss function,we can retain more salient features of dehazed image.Finally,we evaluate our algorithm in both qualitative and quantitative tests on 10 public datasets.In the quantitative experiment,dehazed images are input into SIFT to detect key points and verify the repeatability and robustness of key points.And,NIQE is used to measure whether the image is distorted or not.Finally,we demonstrate the effectiveness of each module by ablation study.The results show that the performance of our algorithm is obviously better than other algorithms in each dataset,which can retain the feature information after dehazing on the real hazy images.