Night-time vehicle model recognition based on domain adaptation

• Published in: Multimedia tools and applications Vol. 83; no. 4; pp. 9577 - 9596
• Main Authors: Yu, Ye, Chen, Weixiao, Chen, Fengxin, Jia, Wei, Lu, Qiang
• Format: Journal Article
• Language: English
• Published: New York Springer US 2024
• Subjects: Computer Communication Networks; Computer Science; Data Structures and Information Theory; Multimedia Information Systems; Special Purpose and Application-Based Systems; Domain adaptation; Generative adversarial networks; Image decomposition; Night-time Vehicle Model Recognition (NVMR)
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-15447-1

Owing to the low brightness, low contrast, and high labeling difficulty of night-time vehicle images, night-time vehicle model recognition (NVMR) faces significant challenges. To address these challenges, this paper proposes the Night-time Vehicle Model Recognition (DA-NVMR) method based on domain adaptation theory and Generative Adversarial Networks (GANs). The proposed method realizes the NVMR based on the VMR model trained on daytime vehicle images. In DA-NVMR, a Domain Adaption Network (DANet) is designed to find the mapping between night-time vehicle images and daytime vehicle images. The DANet consists of two modules: the decomposition module and conversion module. The decomposition module decomposes the vehicle images into reflectance images and illumination images. The conversion module realizes the conversion between the illumination of the night-time and daytime vehicle images. Experiments based on the simulated night-time vehicle dataset and real night-time vehicle dataset confirmed that DA-NVMR can effectively identify night-time vehicle models. Compared with other low-light image enhancement methods and domain adaptation methods, the Top-1 recognition accuracy of the proposed method improved by at least 2 % and 1 % , respectively.