Abstract: To address the problem that existing low-light image enhancement methods primarily focus on perceptual optimization while struggling to simultaneously support downstream high-level vision tasks such as object detection, an unsupervised two-stage method, termed DBDiff, was proposed by integrating diffusion priors with a detection-bridging network so as to improve both perceptual quality and task utility without requiring paired training data or manual detection annotations. In the first stage, exposure-guided pretrained diffusion sampling was employed to progressively restore a natural illumination distribution, while an Adaptive Homomorphic Filtering (AHF) module was introduced to perform frequency modulation during the reverse denoising process, thereby alleviating the attenuation of high-frequency texture details. In the second stage, an unsupervised bridging network composed of an Adaptive Weighted Fusion (AWF) module and a Task Alignment Module (TAM) was constructed. Specifically, cycle-consistency constraints were imposed to preserve content structural consistency, and mixup augmentation, pseudo-label-based self-training, and downstream detection loss were further combined to align the enhanced results with the input distribution preferred by the detector. Experimental results showed that, on the public ExDark and DarkFace datasets, DBDiff achieved Natural Image Quality Evaluator (NIQE) scores of 3.991 and 3.314, Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) scores of 21.796 and 12.053, and Perception-based Image Quality Evaluator (PIQE) scores of 36.150 and 21.262, respectively, yielding the best overall performance in no-reference quality evaluation. In downstream object detection experiments, the mAP50 was improved by 1.8 and 1.7 percentage points over the best competing methods, namely FourierDiff on ExDark and CLIP-LIT on DarkFace, respectively. These results demonstrated that the proposed method could effectively support downstream low-light vision tasks such as object detection while achieving natural exposure and faithful detail preservation.

Key words: low-light image enhancement, diffusion model, unsupervised learning, detection-oriented, domain alignment

摘要： 针对现有低光图像增强方法偏重视觉感知优化、难以兼顾目标检测等下游高级视觉任务性能的问题，提出一种将扩散先验与检测桥接网络相结合的无监督两阶段方法DBDiff，在无需配对训练数据和人工检测标注的条件下，提升感知质量和任务可用性。第一阶段，利用曝光引导的预训练扩散采样逐步恢复自然照明分布，并引入自适应同态滤波（AHF）模块，在反向去噪过程中进行频率调制，以缓解高频纹理细节衰减；第二阶段，构建由自适应加权融合（AWF）模块与任务对齐模块（TAM）组成的无监督桥接网络，通过循环一致性约束保持内容结构一致性，并结合mixup增强、伪标签自训练与下游检测损失，将增强结果对齐至检测器偏好的输入分布。实验结果表明，DBDiff在公开数据集ExDark和DarkFace上的NIQE（Natural Image Quality Evaluator）分别达到3.991和3.314，BRISQUE（Blind/Referenceless Image Spatial Quality Evaluator）分别达到21.796和12.053，PIQE（Perception-based Image Quality Evaluator）分别达到36.150和21.262，无参考评价指标综合表现最优。在下游目标检测实验中，平均精度均值mAP50较ExDark上的最优对比方法FourierDiff和DarkFace上的最优对比方法CLIP-LIT分别提高1.8和1.7个百分点，表明所提方法在实现自然曝光和真实细节的同时，能够有效支撑目标检测等下游低光视觉任务。

关键词: 低光图像增强, 扩散模型, 无监督学习, 检测导向, 域对齐