Research On Single Image Restoration Methods In Foggy And Rainy Environment

The frequent occurrence of fog and rain in daily life has a strong interference effect on the propagation path of light in the atmospheric medium,which seriously affects the imaging effect of visual equipment,resulting in quality degradation of the captured image,such as excessive image noise,reduced sharpness,low color saturation,etc.The subsequent image recognition,image semantic analysis,image data extraction and other work cannot be performed normally.Therefore,it is of great significance and academic value to restore single images in foggy and rainy environments by combining the corresponding image physical models.Aiming at the single image in fog and rain environment,combining with the corresponding image physical model on the basis of deep learning technology,this thesis proposes a method of single foggy image restoration based on region enhancement and a method of single rainy image restoration based on concentration determination to remove the fog noise and rain streak noise in the image respectively,so as to restore the image.The main research contents are as follows:(1)A method of single foggy image restoration based on region enhancement is proposed.At present,the method of image restoration in foggy days cannot be combined with the depth change of image scene to carry out better adaptive processing,resulting in a large amount of fog noise remaining in the restored image.Aiming at this problem,this thesis designs and constructs a propagation map prediction network to regression the mapping relationship between various features and optical path propagation in the image,and obtain the corresponding propagation map.After that,combined with the region depth and characteristics of the foggy image,an adaptive scene division and Enhancement strategy is proposed to adaptively adjust the dehaze intensity according to the depth change of different regions,thereby the fog noise in the image is removed by combining the physical model of foggy image,so as to achieve the goal of restoring the image.(2)A method of single rainy image restoration based on concentration determination is proposed.At present,the method of single rainy image restorationlacks ability of rain streak concentration determination,neglects the information of rain streak concentration in the image,and leads to the problem of rain streak noise remaining in the restored image.Therefore,based on the deep learning technique,this thesis proposes a rain streak concentration determination network to adaptively divide and determine the rain streak concentration in the image,and to output the corresponding rain streak concentration label.At the same time,a rain streak noise prediction network is constructed to extract the physical characteristics of the rain streak,and the feature is fused with the rain streak concentration label to obtain a more complete rain streak noise map.Experiments show that this method has better prediction ability of rain streak noise,and the image is clear and natural in combination with the physical model of rainy image,and has high sharpness.(3)The experimental and results analysis of the image restoration method proposed in this thesis.In order to evaluate the objective experimental results of the image restoration method proposed in this thesis,based on the subjective evaluation,this thesis combines the existing objective evaluation indicators of image quality to establish an image quality evaluation system combining subjective and objective,and using the current mainstream third-party image dataset to compare the experimental results of multiple methods,the experimental results of this method are analyzed and evaluated from the subjective and objective dimensions.The experimental results show that the single image restoration method proposed in this thesis in foggy and rainy environment has good image restoration effect,and have a great improvement in image clarity and color reproduction.