The Research Of Single Image Dehazing Algorithm Based On Deep Learning

Photographs taken in hazy weather often have problems such as blurred images and loss of detail,which reduces the clarity of the image and severely affects the depth of processing of the image in post.The image-defogging algorithm can reduce or even eliminate the effect of haze on the image so that the restored haze-free image is more explicit,accurate,and detailed.This thesis presents an in-depth study of image defogging methods based on deep learning,as following main work:Firstly,to address the problem that the deep learning-based fog removal algorithm is limited by the synthetic data set,which leads to poor scene generalization ability and loss of depth details,this thesis proposes a fog removal method based on an improved feature fusion Cycle GAN.Firstly,an improved residual dense block is introduced in the generator network instead of the convolution module to effectively extract feature information under different perceptual fields in foggy images;Secondly,to address the complexity of haze distribution in natural scenes,a feature fusion structure incorporating an attention mechanism is introduced in the hopping connection of the network to splice and fuse the shallow features of each region with the in-depth features to improve the resolution and detail information of the generated images;Finally,in order to improve the quality of the generated fog-free images,the perceptual loss is introduced to enhance the detail information of the extracted features,making the generated fog-free images more realistic.The experimental results show that the proposed defogging algorithm can have a better defogging effect in natural scenes.Secondly,to address the problem that the convolutional neural network-based image defogging algorithm fails to fully use the fogged image’s feature information,resulting in the blurred texture of the defogged image,this thesis proposes an algorithm for image defogging based on multi-scale residual attention.The proposed algorithm designs a multi-scale residual module to extract and fuse more rich,detailed information.At the same time,a hybrid attention mechanism is introduced in the multi-scale residual module,enabling the defogging network to adjust adaptively to different regional features.In addition,the acceptance domain of the network is expanded by adding residual smoothing expansion convolution after downsampling,which can improve the recovery ability of the network for small objects without increasing the parameters.The experimental results show that the proposed algorithm improves subjective and objective evaluations,and the recovered fog-free images have more detailed and natural textures.Finally,this thesis combines the proposed image-defogging algorithm to design a vehicle recognition system under foggy high-speed surveillance video.The processing time and memory consumption of video defogging is reduced by extracting and defogging the keyframes containing vehicles in the foggy video.The vehicle recognition system consists of a critical frame extraction module,a haze detection module,an image defogging module,and a vehicle detection module,which can accurately recognize vehicles driving in high-speed surveillance video more efficiently.