Nighttime Semantic Segmentation with Attention and Low-Light Enhancement
With the development of deep learning technology and the improvements in computing power,semantic segmentation of natural scene images captured during the day shows promising results.However,in nighttime image semantic segmentation tasks,models trained on daytime data often fail to deliver satisfactory performance due to challenges such as imbalanced exposure and a lack of labeled data.To address these challenges,a new unsupervised nighttime image semantic segmentation network called AI-USeg is proposed.First,a lightweight Self-Calibrating Illumination(SCI)network is used to enhance nighttime images,thereby mitigating the impact of lighting variations on subsequent semantic segmentation networks.Next,a Domain Adaptation(DA)method is introduced to transition the model from Cityscapes containing a large amount of labeled data to Dark Zurich-D,addressing the lack of labeled data.Subsequently,AI-USeg introduces a Squeeze-and-Excitation Network(SENet)into the discriminator,built upon a Fully Convolutional Network(FCN).This adaptation facilitates the adjustment of image features in low-light nighttime settings through adversarial learning in the output space,ultimately improving the performance of semantic segmentation tasks for nighttime images.The experiment used two sets of 2 416 day and night image pairs sourced from Cityscapes and Dark Zurich-train for unsupervised training.The results show that AI-USeg achieved Mean Intersection over Union(mIoU)values of 47.9%and 51.5%on the Dark Zurich-test and Nighttime Driving-test datasets,respectively.These values were 5.4 and 2.1 percentage points higher than those obtained using the MGCDA method.In conclusion,AI-USeg displayed stronger adaptability to nighttime image features and higher robustness than previous segmentation models,providing an effective solution for image segmentation tasks in nighttime scenes.
deep learningsemantic segmentationautonomous drivinglow-light image enhancementattention mechanism
国家科技期刊平台
NSTL
万方数据
