Context-guided adaptive feature learning for robust low-light object detection.

Abstract

Low-light object detection is a challenging task in computer vision and visual computing, with wide applications in autonomous systems, underwater monitoring, and nighttime surveillance. Images captured under insufficient illumination usually suffer from low contrast, noise amplification, and degraded texture details, which reduce detection reliability. This work proposes CMSA-Net, a context-guided multi-scale adaptive network for low-light object detection. CMSA-Net integrates an adaptive context-split backbone, input-adaptive enhanced attention, and multi-scale adaptive fusion to improve feature representation under illumination degradation. On ExDark and RUOD, CMSA-Net achieves 71.89% and 88.10% mAP@0.5, respectively, while maintaining a lightweight model size and real-time inference speed. A self-built controlled darkroom imaging platform is further used to analyze detector behavior under different illuminance, exposure time, and ISO settings, showing that CMSA-Net provides stable gains under low and mid-to-low effective exposure conditions. These results demonstrate the effectiveness of CMSA-Net for robust object detection in low-light imaging scenarios.