Image Dehaze And Application Based On Deep Learning

In haze weather,due to suspended fine particles in the atmosphere affect the propagation path of light,the video or image collected by the imaging system becomes degradation such as contrast and sharpness decrease.Not only does it seriously affect the visual effect of the image,but also limits the application value of digital images in various fields.Therefore,it is of great research significance to use image dehaze technology to reduce or clear the impact of haze on the image and restore the original details of the image.In recent years,image dehaze methods based on deep learning have become a research hotspot in the field of computer vision and have made significant progress.However,most algorithms rely on synthetic data for training,which causes the effect of dehazing in real scenes is not very ideal.In addition,there are still great challenges in the balance of accuracy and efficiency of the existing dehazing algorithms.Aiming at existing problems of the current methods,this paper has done substantial research work in the direction of single image dehazing based on deep learning technology,main contents of this article are as follows:1.This thesis proposes an enhanced image dehaze algorithm based on cycle consistent adversarial network.The algorithm is an end-to-end network structure,which does not rely on pairs of synthetic data to learn fogged image characteristics.This method improves the cycle consistent adversarial network with combining the dense connection,residual structure,and resize-convolution to enhance the dehazing ability of the model.Aiming at the phenomenon of loss of image content and color distortion generated by the network,a few joint loss functions are added to improve the accuracy of training and ensure the visual quality of the generated images.Experimental results show that the network has strong generalization and can achieve good results in real outdoor scenes.2.This thesis proposes a novel progressive defogging network based on perceptual fusion mechanism(MPD-Net)to improve the efficiency and accuracy of image dehaze.This method decomposes the task of degraded image recovery into multi-stage subtasks,and uses lightweight subnetwork chunk learning semantic information of different regions of the feature map to ensure image dehaze efficiency.On this basis,the cross-stage perception fusion module(PFM)is introduced to cascade the features extracted by each stage which is based on the attention mechanism and guided filtering,and can be fused with adaptive perceptual semantic features without losing image edges and texture details.According to experimental results,the method suggested in this research offers superior accuracy and real-time performance on complicated outdoor photos than the existing mainstream end-to-end fog removal approach,with PSNR improving by 0.93 d B compared with the best results available and processing a single image in only 72 ms.3.This thesis designs an image dehazing system for monitoring scenes to solve the problem that the visibility of captured images under haze weather which affects the normal work of various fields.The system consists of image acquisition,haze detection,and dehazing modules that enable fog removal both high-quality and high efficiency.The system can be applied to the field of intelligent security to complete the pre-processing of the video surveillance system,and solve the follow-up face recognition and target tracking tasks in the public security management problem,which effectively improve the work efficiency of the supervisors.It is foreseeable that this paper system can also help other areas.