Low-light conditions pose significant challenges for the extraction of visual features essential for downstream tasks. Previous research has attempted to enhance image representations either by correlating visual quality with machine perception or by designing illumination degradation transformation methods that necessitate pre-training on synthetic datasets. However, we propose that optimizing the enhanced image representation in relation to the specific loss functions of downstream tasks can yield more expressive representations. In this study, we introduce a novel network, LPEnHancer, which consists of two key modules. By utilizing the Laplacian pyramid, the input low-light images are decomposed into multiple scales, where the Detail Mining Module is used to excavate edge/texture features and enhance local features at each scale, and the Low Frequency Enhancement Filter is employed to suppress high-frequency noise. LPEnHancer is a versatile plug-and-play module that can be seamlessly integrated into any low-light vision pipeline. Our extensive experimental results demonstrate that the enhanced representations generated by LPEnHancer significantly and consistently improve performance across various low-light vision tasks, including dark object detection (3.0 mAP improvement on ExDark), dark face detection (2.0 mAP improvement on DARK FACE), nighttime semantic segmentation (0.78 mIoU improvement on ACDC), and dark instance segmentation (1.7 Bbox_mAP and 0.9 Segm_mAP improvements on LIS).

low-light object detection, plug-and-play, pyramid enhancement, multi-scale