Deep Learning Based De-raining For Images In Natural Scenes

Various monitoring systems have been widely used in the fields of intelligent transportation,public safety and environmental protection to perform tasks such as target recognition and tracking.In outdoor scenarios,especially when facing severe weather conditions such as rain and fog,many intelligent monitoring systems begin to function abnormally,owing to the degradation of image quality.Therefore,how to restrain the influence of adverse weather,improve the image quality,and expand the scope of the existing visual system to enable the all-weather monitoring is an urgent and important problem to be solved,which has theoretical significance and use value.Although the existing image clearing methods have achieved some fruitful results,in the face of complex rain and fog weather,these algorithms still cannot meet the requirements for image repair quality and processing speed,especially in local area smoothing and detail texture protection On the issue,more advanced theories and technical solutions are urgently required.In addition,for the deep learning rain removal method,the synergy between the loss of multiple tasks and the robustness of the loss function are research directions that deserve attention.This article starts with the above problems and studies a single image de-raining method based on deep learning.In response to these problems,this paper focues on single image de-raining via deep learning and achieves the following results:(1)Propose a multi-constraint target that describes image quality differently from existing methods.By quantitatively describing the image content,edges,and local texture self-similarity and designing the corresponding constraint loss,the model solution space during training is enhanced Tightness limitation,thereby reducing the degree of fuzzy distortion of the predicted image;(2)Two dynamic weighting algorithms are proposed for multi-task coefficient setting,respectively,to solve the problem of hyperparameter setting from the perspective of gradient balance and loss balance,and to further tap the potential of the model;(3)Propose a novel loss function with both edge-preserving and smoothing characteristics.By changing the statistical mean value characteristics of the original loss function without distinction,the model's ability to protect background details and textures is improved;(4)Before the object detection,pre-process the rainy image to prove the actual value of this study.