Low-Light Object Detection via Illumination-Invariant Feature Fusion and Enhancement-Detection Joint Module

Low-light environments pose severe challenges to object detection systems due to low contrast, heavy noise, and weak or uneven illumination. Traditional detectors often suffer from missed detections or false positives under such conditions. In this paper, we propose a novel framework, IIF-Det (Illumination-Invariant Feature Detector), which integrates illumination-invariant feature learning, adaptive enhancement, and a joint enhancement-detection fusion module in an end-to-end training manner. Specifically, we design an Illumination-Invariant Feature Module (IIFM) based on a simplified Lambertian model to extract features robust to lighting changes. In parallel, an Adaptive Enhancement Branch (AEB) refines local contrast and suppresses noise. Then, a Fusion and Detection Module (FDM) merges the two streams and performs detection, allowing mutual guidance between enhancement and detection. We conduct experiments on multiple low-light and normal-light datasets (e.g. ExDark, NOD, COCO-d). Our method achieves state-of-the-art performance under low illumination, outperforming baseline detectors by significant margins in mAP, and maintains competitive performance under normal lighting. Ablation studies validate the effectiveness of each module.