Research On Method Of Video/Image Dehazing And Deraining Based On DEEP RESNET

Image quality enhancement is an indispensable preprocessing process in computer vision tasks such as object recognition and detection,three-dimensional reconstruction,motion and tracking.The goal of the video/image deraining and dehazing algorithm is to restore the image contaminated by rain and fog into a clear image,so as to achieve the purpose of image enhancement.The existing techniques for the algorithm of videos and images dehazing and deraining only perform well in one of dehazing task and deraining task,with a certain singleness.This paper studies an algorithm that can simultaneously complete the tasks of rain removal and fog removal,and proposes a videos and images deraining and dehazingl based on deep residual network.The main work is as follows:1)Aiming at the problems of low real-time performance,complex parameters,and low rain and fog removal efficiency caused by the current dehazing and rain removal algorithm due to the deep network model,this paper proposes removal network model of videos and images deraining and dehazing based on an improved deep residual network,which combines the attention mechanism and the deep residual network,and the attention mechanism adaptively learns rain and fog features from the two dimensions of channel and space,so that the network model has been simplified.The algorithm proposed in this paper can not only remove raindrops and rain streaks in all directions,but also remove fog of different concentrations,with high real-time performance,and the processing running time for a single image is within 0.1 s;The error value of the rain and fog removal network model of the poor network converges around 0.004,which has high accuracy and has wide application value.2)Aiming at the problem that the image processing results of some algorithms of rain and fog removal are too smooth and lose background details,this paper solves the problem from two perspectives: the detail preservation module and the loss function.First of all,this paper adds a detail preservation module to the network model,which is composed of a discriminator in the generative adversarial network,which can learn the difference between the output image and the clean background image,so as to ensure that the image details are not lost.Secondly,some algorithms use the mean square error function as the loss function,which will cause the effect to be over-smoothing.In this paper,the structural similarity function and the adversarial loss function are added on the basis of the mean square error function,which better solves the problem of over-smoothing.Compared with other loss functions,the PNSR value of the loss function method used in this paper reaches 27.6 d B,and it reaches 0.936 on the SSIM index.3)This paper selects the existing dehazing data set RESIDE,rain pattern data set Rain12000 and Rain100 L data set,raindrop data set Raindrop,and collects a large number of real-world dynamic scene rain and fog videos and images to train the network model,so that the network model in this paper can remove It is more robust to fog and rain of different concentrations.