Research On Rain And Haze Image Enhancement Based On Deep Learning

With the continuous development of traffic information collection technology,it is a challenge and opportunity for the traffic industry to mine and extract high-quality information from massive information.In the process of traffic image collection,weather will affect the collected images.Rain and haze are common weather.These two kinds of weather can degrade images collected by image acquisition equipment.Image enhancement in rainy and haze weather is an ill posed problem.The solution to this problem can be divided into two parts: the traditional method and the method using deep learning.The traditional method usually uses prior knowledge combined with the optical model of the rain and haze image to enhance the degraded image.The image enhancement method based on deep learning is usually the application of the image enhancement method in the rain and haze image.In this paper,we use the generation countermeasure network model in deep learning to enhance the image of rain and haze.The main contents of this paper are as follows:The optical model of rain fog image is analyzed from the point of view of rain and haze image imaging.Aiming at three different reasons for the quality degradation of rain and haze image,the model is built respectively.The mathematical expression of the optical model of image quality degradation under the condition of rain and haze is obtained by combining the three models.The framework of generation adversarial network is constructed based on the research of neural network which is commonly used in image processing.By studying convolution neural network,residual neural network and generation adversarial network,this paper proposes a generation adversarial network model framework of rain and haze image enhancement,which includes expandable auxiliary generation network,defines target loss function,and introduces cycle consistency loss and cycle sense consistency loss.On the basis of extending the auxiliary generation network,the performance of the proposed network model is significantly improved by increasing the constraint loss of the optical model.The generation adversarial network including the scalable auxiliary generation network is verified combined with the actual data.Through selecting the commonly used open data set of rain and haze enhancement to train and test the network,and using the actual collected traffic scene pictures as the test set to evaluate the network model.After the calculation of PSNR and SSIM,images based on the model have a good performance and meet the expected requirements.The network is verified by combined with the actual data.By selecting the widely used rain image data set and haze image data set to train the generated network,the generated network in the training results is taken as the evaluation target.After evaluation,the performance of the test set is better than that of the early image enhancement method,which is close to the advanced level of the industry and meets the expected requirements.