| The propagation of light in hazy weather is hindered by particles suspended in the atmosphere,leading to a drastic decline in the quality of images or videos acquired by the imaging system.This is mainly reflected in reduced visibility,increased brightness in foggy regions,and a lack of detail.The development of advanced visual tasks such as target tracking and automatic driving is hindered by the visual effects of videos,as well as the transportation of people due to haze,which can lead to serious traffic accidents.Consequently,dehazing technology has significant research implications,This paper will explore these issues in greater detail.First of all,to address issues such as the loss of detail information in foggy images and incomplete defogging after defogging,this paper developed a mesh network to extract feature information from images with fog.This is in contrast to the current video dehazing methods which are mainly based on single frame image dehazing,as they are unable to handle continuous video dehazing.Multiple up-sampling and down-sampling are used to complete the information exchange at different scales,which can retain more details.The model presented a channel attention system to join together multiple features.Every channel of the feature map was given the same significance before the attention mechanisms input,and it could augment the channels weight with more fog features,thus lessening the need to calculate insignificant data and thus hastening its operation.Short-circuit connections in residual-dense blocks allow rich low-frequency information to propagate backward directly through identity mapping,ensuring the flow of information without generating additional parameters and enabling efficient use of relevant features.Secondly,the algorithm proposed in this paper considers the root cause of image imaging with fog in foggy environment.Most of the two-stage dehazing methods based on deep learning use neural networks to estimate transmission from foggy images,then use the empirical method to estimate atmospheric light value,and finally use the model formula to output non-fog images.Such a process will lead to a suboptimal solution.Errors will accumulate or even magnify,resulting in the loss of image details.Therefore,in this paper,transmission and global atmospheric light two parameters are unified into one parameter,which can change with the change of fog image,and minimize the reconstruction error between the fog removal image and the original clear image,the obtained fog removal image will retain more image details.Thirdly,most deep learn-based dehazing methods,however,only utilize images without fog to oversee model training,disregarding the data of foggy images.Therefore,contrast regularization model training employs a loss function based on contrast learning,with negative samples as the basis information is used.This document employs two algorithms,one for indoor datasets and the other for outdoor datasets and dehaze images,to conduct experimental tests.The network is ascendency to the contrast methods in terms of PSNR and SSIM evaluation indexes and subjective visual effect,thereby enhancing the visual effects of dehazing images.Finally,this paper designs an intelligent image dehazing system based on Py Qt5,which integrates the algorithm proposed in this paper and the mainstream dehazing algorithm,and subjectively evaluates different dehazing algorithms by image generation quality.In fog and haze weather,the system can be utilized to resolve the difficulty of distinguishing traffic conditions,signs,and vehicle license plates,thereby enhancing the work productivity of personnel and diminishing the frequency of traffic mishaps. |
