光照感知特征增强的低照度船舶目标检测方法

Illumination-Aware Feature Enhancement Method for Low-Illumination Ship Target Detection

摘要: 针对低照度条件下可见光船舶图像质量退化导致检测精度下降的问题，本文提出一种光照感知特征增强方法IAFE (Illumination-Aware Feature Enhancement)。与传统的“先增强图像、后检测”两阶段方案不同，IAFE直接在骨干网络的特征空间中根据局部光照强度对特征进行自适应调制——暗区域的船舶特征被选择性放大，亮区域的特征保持不变。IAFE以YOLOv8s为基线模型，在骨干网络P3/P4/P5三个尺度的输出层嵌入轻量级光照感知调制模块，每个模块仅增加约0.01 M参数。在合成低照度SSDD数据集上的实验表明：IAFE将严格指标mAP@0.5:0.95从70.49%提升至70.85% (+0.36个百分点)，验证了特征空间光照感知增强优于图像空间增强的技术路线。消融实验则验证了在三个特征层级上均部署IAFE的必要性。与Zero-DCE、Retinex-Net、KinD等经典低光照增强方法的对比表明，所有图像空间增强方法在严格检测指标上均出现退化(−0.11%至−0.60%)，仅IAFE获得正向增益。针对mAP@0.5下降现象的系统错误分析揭示了IAFE对检测器精度–召回特性的重新校准机制。在HRSID数据集上，通过将IAFE扩展至浅层高分辨率特征(P2)，增益从+0.02%提升至+0.25%，验证了高分辨率特征调制对改善复杂场景性能的有效性。

Abstract: Aiming at the detection accuracy degradation caused by quality deterioration of visible light ship images under low-illumination conditions, this paper proposes an Illumination-Aware Feature Enhancement (IAFE) method. Different from the conventional two-stage paradigm of “image enhancement followed by target detection”, IAFE adaptively modulates features in the feature space of the backbone network according to local illumination intensity: ship features in dark regions are selectively amplified, while features in bright regions remain unchanged. Taking YOLOv8s as the baseline model, IAFE embeds lightweight illumination-aware modulation modules into the output layers of three scales (P3/P4/P5) of the backbone network, with each module adding only approximately 0.01M parameters. Experiments on the synthetic low-illumination SSDD dataset demonstrate that IAFE improves the strict metric mAP@0.5:0.95 from 70.49% to 70.85% (an increase of 0.36 percentage points), which verifies that the technical route of illumination-aware enhancement in feature space outperforms image-space enhancement. Ablation experiments validate the necessity of deploying IAFE on all three feature levels. Comparisons with classic low-light enhancement methods including Zero-DCE, Retinex-Net and KinD reveal that all image-space enhancement methods suffer performance degradation on the strict detection metric (ranging from −0.11% to −0.60%), while only IAFE achieves positive performance gains. Systematic error analysis for the decline of mAP@0.5 reveals the recalibration mechanism of IAFE for the precision-recall characteristics of the detector. On the HRSID dataset, extending IAFE to shallow high-resolution features (P2) raises the performance gain from +0.02% to +0.25%, which verifies the effectiveness of high-resolution feature modulation for boosting model performance in complex scenarios.