Beyond Unpaired Degradation Learning for Low-Light Image Enhancement in ITS
IEEE Transactions on Intelligent Transportation Systems · 2026-11 · 27 (11) · 14364-14381
DOI: 10.1109/tits.2026.3724322 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Intelligent Transportation Systems, 2026-11. Verfasst von Changhui Hu und 5 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Natural Science Foundation of Nanjing University of Posts and Telecommunications.
Abstract (Originalsprache)
This paper proposes a beyond unpaired degradation learning framework (BUDL) for low-light image enhancement (LLIE) in intelligent transportation systems (ITS). BUDL consists of two phases, 1) unpaired degradation learning via cycle degradation learning network with multiple self-attentions (DLMS), and 2) joint degradations for synthesis night traffic monitoring (NTM) images. On the one hand, as multiple self-attentions can learn different long-range dependences, we believe that the transformer network with multiple self-attentions can learn multiple degradations for complex degradation learning task, and propose two novel self-attentions, i.e., cross-channel row/column self-attention (CRS/CCS). Based on this innovation, DLMS is constructed by three different self-attentions, i.e., CRS, CCS, and intra-channel block self-attention (IBS), which can efficiently mitigate bright spots of degradation low-light images generated by cycle DLMS. On the other hand, motivated by the fact that cycle DLMS can learn abstract shared degradations from unpaired NTM images, rather than their specific detail degradations, we propose a multiple mathematical degradation (MMD) model to synthesize specific detail degradations, including blurring, down-sampling, noise, compress, and zero-element pixels. Experiments show that BUDL can generate high-quality synthesis NTM images, and makes existing supervised learning LLIE methods be able to enhance real NTM images. Our source codes and pre-trained models are available at https://github.com/NJUPT-IPR-ChenTS/BUDL
Förderung
- National Natural Science Foundation of China
- Natural Science Foundation of Nanjing University of Posts and Telecommunications
Zitieren
APA
Hu, C., Chen, T., Han, Z., Guo, Y., Jing, X., & Liu, P. (2026). Beyond Unpaired Degradation Learning for Low-Light Image Enhancement in ITS. IEEE Transactions on Intelligent Transportation Systems, 27(11), 14364-14381. https://doi.org/10.1109/tits.2026.3724322
Vancouver
Hu C, Chen T, Han Z, Guo Y, Jing X, Liu P. Beyond Unpaired Degradation Learning for Low-Light Image Enhancement in ITS. IEEE Transactions on Intelligent Transportation Systems. 2026;27(11):14364-14381. doi:10.1109/tits.2026.3724322
BibTeX
@article{Hu2026,
author = {Hu, Changhui and Chen, Tiesheng and Han, Zhongfa and Guo, Yanyong and Jing, Xiaoyuan and Liu, Pan},
title = {Beyond Unpaired Degradation Learning for Low-Light Image Enhancement in ITS},
journal = {IEEE Transactions on Intelligent Transportation Systems},
year = {2026},
volume = {27},
number = {11},
pages = {14364--14381},
doi = {10.1109/tits.2026.3724322}
}
EUVIMED