Coal mine safety relies heavily on video surveillance, yet low-light environments and limited computational resources pose serious challenges to reliable object detection. This study aims to address these issues by proposing a two-stage framework that integrates unsupervised low-light image enhancement with lightweight object detection. First, we construct and release the Coal Mine Low-Light Image Dataset (CMLOL), a publicly available dataset designed for underground coal mine monitoring. Based on CMLOL, we design MCTE-GAN, an enhancement model that introduces matrixed color transformation into the generator to improve global brightness while preserving structural details. Second, we develop YOLO-PKD, a lightweight detector that combines structured channel pruning with response-based knowledge distillation, significantly reducing model complexity without compromising accuracy. Experiments show that MCTE-GAN effectively enhances image quality under poor lighting, while YOLO-PKD achieves real-time detection at 139.1 FPS with only 2.5 M parameters and 4.2MB model size, nearly matching the accuracy of larger models. Overall, our method provides an efficient and scalable solution for intelligent coal mine monitoring, demonstrating strong adaptability for deployment in resource-constrained environments. • Proposed a public unpaired low-light image dataset for coal mine monitoring. • Developed an unsupervised low-light enhancement method via matrixed color transform. • Designed a lightweight detector via structured channel pruning and knowledge distillation. • Real-time inference and stable performance in multi-platform deployment validate practicality for coal mines.
Li et al. (Tue,) studied this question.