| It is well known that foggy environments have a significant impact on the visual perception capabilities of images.The emergence of fog creates a blurring effect on the visual system,making the objects in the scene less clear,causing loss or significant reduction of visual features,presenting significant challenges for the perceptual model’s interpretation and understanding of the image.Therefore,image defogging under foggy conditions is a key step in improving the perception capabilities of machine vision systems.At the same time,visual perception tasks such as semantic segmentation are important research topics in the field of autonomous driving and still face challenges from various weather scenarios such as rain,snow,and fog.Especially,obtaining annotation data for autonomous driving under real foggy conditions is even more difficult,since training samples are extremely scarce.Traditional visual perception methods,including most domain adaptation methods,either focus on a single weather type or assume that data can be obtained equivalently under different weather scenarios.Therefore,how to utilize weather scenario data that is relatively easy to obtain,such as rain,snow,and cloudy,reasonably mine the correlation of data distribution between different domains,and explore the cross-domain transfer of semantic segmentation under normal weather to scarce scenarios like foggy weather,has important theoretical and practical significance.This article mainly uses the foggy dataset,game dataset,real driving scenario dataset BDD100 k,and the foggy dataset Foggy-City Scapes as research objects.On one hand,it enhances the performance of the defogging network and improves the quality of defogged images by introducing a new numeric format inspired neural network module.On the other hand,as for road scene semantic segmentation,it combines attention mechanisms to explore the cross-domain enhancement of domain-adaptive neural networks,achieving the better transfer of semantic segmentation networks from normal weather to foggy and other weather conditions.The main research content is as follows:(1)Defogging methods based on deep neural networks are increasingly attracting attention,where the key network module is the important part of the performance of deep learning models.For this reason,this paper constructs a neural network module Prog Net,inspired by numeric formats,which enhances the interpretability of the network module by connecting the network structure with the forward Euler progressive multi-step numeric format.Taking a further step,Prog Net is introduced into the two recent typical deep defogging models PFF-Net and AOD-Net,constructing PFF-Prog Net and AOD-Prog Net.In addition,to make up for the limitations of the classic atmospheric scattering model in the defogging network,this paper further draws on the extended defogging model of PFF-Net.Experimental results on public datasets verify the effectiveness of the new Prog Net module for defogging various foggy scenes and enhance the defogging performance of PFF-Net and AOD-Net.This paper conducts dehazing experiments on the RESIDE and NTIRE2018 datasets to evaluate the performance improvement achieved by the Prog Net module.The experimental results on the RESIDE dataset demonstrate that AOD-Prog Net dehazing yields a PSNR improvement of 2.1 and an SSIM improvement of 0.1566,while PFF-Prog Net dehazing achieves a PSNR improvement of 2.63 and an SSIM improvement of 0.0013.Similarly,on the NTIRE2018 dataset,AOD-Prog Net dehazing exhibits a PSNR improvement of0.64 and an SSIM improvement of 0.0647,whereas PFF-Prog Net dehazing achieves a PSNR improvement of 1.91 and an SSIM improvement of 0.0202.The experimental results confirm that the proposed Prog Net module enhances the performance of dehazing networks.(2)In response to the practical difficulties of not being able to obtain equivalent amount of data under different weather scenarios,this paper proposes a fog-free cross-domain semantic segmentation method SECB-OCDA.This method builds on the baseline network OCDA and selects semantic segmentation for domain adaptation.Firstly,it combines the semantic segmentation network and SE-Net visual channel attention mechanism to enhance feature extraction capabilities.Then,the method combines the domain encoding network and CBAM mixed attention mechanism and proposes a new domain encoder structure,thereby enhancing domain feature extraction capabilities and achieving better network transfer capabilities.This paper uses GTA-5 as the source domain,various weather data of BDD100 k as the target domain and Foggy-Cityscapes foggy dataset as an open domain that does not participate in training.The experiments prove that SECB-OCDA improved OCDA’s cross-domain semantic segmentation performance in both non-foggy and foggy scenes.Besides,the cascade combination of SECB-OCDA and PFF-Prog Net achieves further enhancement of cross-domain semantic segmentation performance. |
