| 基于多尺度高低频融合增强恶劣环境下的目标检测 |
投稿时间：2025-11-12 修订日期：2025-12-24  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 4 |
| 全文下载次数: 0 |
|
|
| 中文摘要:在弱光与雾霾环境下，目标检测因对比度降低、噪声增加与结构模糊而面临挑战。本文提出HLF-YOLO，一种基于YOLOv8的改进框架，通过嵌入高低频融合模块实现多尺度特征提取：高频增强滤波器采用多尺寸卷积核捕捉纹理与结构；低频增强滤波器结合混合池化与CBAM注意力模块保留关键低频信息，并通过自适应加权融合机制整合高低频特征。在ExDark与RTTS数据集上的实验表明，HLF-YOLO在弱光环境下达到62.2% mAP（提升3.0%），雾霾环境下达到77.7% mAP（提升1.9%），在精度与速度上均优于现有先进模型，提升了视觉系统在恶劣环境下的可靠性。 |
| 中文关键词:复杂环境下的目标检测 高频纹理细节 低频结构信息 多尺度增强 |
| |
| EnhancingObjectDetectionRobustnessinAdverseConditionsviaMulti-ScaleHigh-LowFrequencyFusion |
|
|
| Abstract:Under low-light and hazy conditions, object detection faces challenges due to reduced contrast, increased noise, and structural blur. This paper proposes HLF-YOLO, an improved framework based on YOLOv8, which incorporates a high-low frequency fusion module for multi-scale feature extraction: a high-frequency enhancement filter with multi-scale convolution kernels captures texture and structure, while a low-frequency enhancement filter combines hybrid pooling and the CBAM attention module to retain key low-frequency information. Features are integrated via an adaptive weighted fusion mechanism. Experiments on the ExDark and RTTS datasets show that HLF-YOLO achieves 62.2% mAP in low-light conditions (an improvement of 3.0%) and 77.7% mAP in hazy environments (an improvement of 1.9%), outperforming existing advanced models in both accuracy and speed, thereby enhancing the reliability of vision systems in harsh environments. |
| keywords:Object Detection in Challenging Conditions, High-frequency Texture Details, Low-frequency Structural Information, Multi-scale Enhancement |
| 查看全文 查看/发表评论 下载pdf阅读器 |
位访问者,今日访问
次