This article presents a targeted approach to the long-standing problem of nighttime object detection, which is critical for pedestrian safety and autonomous vehicle navigation but is hampered by dim and uneven lighting. The authors, Kuo-Feng Hung and Kang-Ping Lin, describe experiments published in the Journal of the Chinese Institute of Engineers (volume 48, issue 5, pages 655-668). The manuscript was received 25 April 2024, accepted 26 March 2025 and published online 23 April 2025. Funding was provided by the National Science and Technology Council [MOST-111-2622-E-033-002].
The study proposes an innovative pipeline that leverages in-context learning (ICL) and knowledge-based chain-of-thought (KCoT) prompting to steer large language models (LLMs). The LLMs estimate ambient lux levels from contextual cues and recommend optimal gamma correction values for image preprocessing. Those recommendations are applied to a pre-trained YOLO object detector. The key result reported in the abstract is a substantial increase in object detection accuracy, from 81.95% to 91.68%, achieved without collecting or training on massive new datasets. The method therefore focuses on model-guided preprocessing rather than heavy data augmentation or re-training.
The authors conclude that LLMs can play a practical role in refining Artificial Intelligence-driven object detection under low-light conditions, effectively bridging reasoning and image preprocessing through prompting strategies. The paper frames these findings as a foundation for future interdisciplinary work that combines language models and computer vision. The disclosure statement reports no potential conflict of interest. Supplemental materials, figures, and references are available through the article’s DOI: https://doi.org/10.1080/02533839.2025.2491425.
