GAN-Based Controllable Image Data Augmentation in Low-Visibility Conditions for Improved Roadside Traffic Perception
Kong Li, Zhe Dai, Xuan Wang, Yongchao Song, Gwanggil Jeon
Abstract
Ensuring the accuracy of visual perception systems is crucial for the construction of intelligent transportation systems. However, in low-visibility traffic scenarios, the scarcity of data and the high cost of annotation make acquiring high-quality image data a significant challenge. Existing data augmentation methods often struggle to meet the requirements of both realism and diversity when generating the needed data. To address this issue, this study proposes a controllable data augmentation technique. Firstly, we improved the architecture and basic units of the Pix2Pix generator network to optimize the detail texture of generated images. Secondly, we introduced adaptive alpha channel technology to make the blending of generated images with original images more natural. Finally, by adjusting the alpha channel, we can precisely control the degree of fusion to produce augmented data that better meets actual needs. Experiments on the proposed surveillance view nighttime image dataset (SVNTI) and NuScenes dataset demonstrate that our method significantly enhances detection performance in nighttime scenarios, validating the effectiveness and potential of controllable data augmentation technology.