Research On Image Dehazing Algorithm Based On Deep Learning

Nowadays,images and videos are the most important source of information as people get information,and good image quality directly determines the information they can obtain.However,in haze weather,a large amount of fine particles are suspended in the atmosphere.In the process of spreading,it interacts with the light,and the light scattered,so that the video or images collected by the imaging system show low contrast,loss of details,and sharpness,which not only seriously affects the visual effect of the image,but also cause difficulties for the computer system to process the image in the next step.Therefore,defogging hazed images to obtain high-quality and clear images has become an important practical need in daily production life.In this paper,based on existing deep learning defogging methods,a comprehensive study of image-defogging algorithms is carried out to address the current problems,which are summarized as follows:1.Proposed an end-to-end image based on RES2 NET and pyramid pooling.In this method,the context feature is extracted by using the Res2 NET module,and the pyramid pool module is used to fuse the feature information of different scales.This model also has the following improvements in the input information and training objectives: the input image of the network is not only the haze image but the edge information is extracted and combined with the haze image into the network;the training objective is also different from other models,the objective of the network training is the residual difference between the clear image and the haze image.This paper uses a RESIDE dataset to train and test the proposed models to get a better network model.The results show that the model has achieved good results in the subjective and objective evaluations,which greatly improves the problem of color distortion and incomplete dehazing of images after de-fog.2.A two-stage network dehazing algorithm based on a bilateral threshold fusion network is proposed.The above model is used as a detailed feature network,and on top of this,a two-stage network dehazing model is formed by adding an overall feature extraction network.The overall and detailed features are fused using a feature fusion network to obtain high-quality defogged images.The two-stage network consists of an overall feature extraction network and a detail feature extraction network: the overall feature extraction network uses the attention U-Net to extract the overall feature details of the network;the detail feature extraction network first obtains the feature map by the encoder,then performs feature extraction by three Res2 Net modules,and finally up samples the feature map to obtain a feature map with rich detail features.Then the two feature maps are input to the feature fusion network-bilateral threshold fusion network for fusion,and finally,the clear image is output.The results of subjective and objective tests on the RESIDE dataset show that the proposed method is more effective than the previous defogging methods and has better defogging effects.