| 基于增强型特征交互的夜间双模态目标检测网络 |
投稿时间：2026-05-11 修订日期：2026-07-09  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 28 |
| 全文下载次数: 0 |
|
|
| 中文摘要:针对夜间低光照下单模态特征提取能力受限、双模态融合算法在特征交互深度与边界重建方面存在不足的问题，提出了基于增强型特征交互的夜间双模态目标检测网络。该模型基于YOLOv11框架，构建了非对称双分支特征引导主干ADFNet，其通过像素注意力引导融合模块PagFM实现可见光与红外特征的深度交互，解决模态信息淹没问题。在颈部引入DySample恢复小目标边界信息，并利用DualConv对C3k2模块进行全局性改进，缓解特征提取与融合中的语义失配问题。在M3FD数据集上的实验表明，该模型mAP50和mAP50?95分别达到77.5%和52.4%，显著优于主流算法，有效提升了夜间场景下的目标检测性能。 |
| 中文关键词:目标检测 多模态融合 特征交互引导 红外图像 |
| |
| A Nighttime Dual-Modal Object Detection Network Based on Enhanced Feature Interaction |
|
|
| Abstract:To address the issues of limited single-modality feature extraction capability in low-light nighttime environments and insufficient feature interaction depth and boundary reconstruction in existing dual-modal fusion algorithms, a nighttime dual-modal object detection network based on enhanced feature interaction is proposed. Based on the YOLOv11 framework, the model constructs an asymmetric dual-branch feature-guided backbone ADFNet, achieving deep interaction between visible and infrared features through the Pixel-attention-guided Fusion Module PagFM to resolve information drowning. DySample is introduced in the neck to recover small target boundary information, and DualConv is utilized to globally improve the C3k2 module, alleviating semantic mismatch in feature extraction and fusion. Experimental results on the M3FD dataset show that the model achieves 77.5% in mAP50 and 52.4% in mAP50?95, significantly outperforming mainstream algorithms and effectively enhancing object detection performance in nighttime scenarios. |
| keywords:object detection multimodal fusion feature interaction guidance infrared images |
| 查看全文 查看/发表评论 下载pdf阅读器 |
