| 摘要: |
| 多波段图像目标检测识别是重要的多模态基础任务之一，旨在通过不同传感器的成像特性补充完善目标特征来提高目标感知效果，对提升交通、医疗、军事等领域智能化程度具有重要的现实意义。针对目标检测识别中广泛存在的因光照、遮挡、复杂背景导致的检测识别效果不佳的问题，提出一种动态自适应聚合的可见光红外图像目标检测识别方法。通过设计通道注意力混合和二次动态权重连接的动态聚合结构以及语义空间信息交互的多路径扩展Neck结构，充分挖掘多波段图像的互补性，来提升困难场景下多波段图像目标融合检测的平均准确率。经公开数据集测试，相较于不采用动态聚合结构和多路径扩展Neck结构的对比模型，该方法的平均准确率(mean Average Precision,mAP)提高4个百分点以上。 |
| 关键词: 多波段图像 融合识别 动态聚合 深度学习 |
| DOI：10.20079/j.issn.1001-893x.240822002 |
|
| 基金项目:国家自然科学基金青年科学基金（62303433） |
|
| A Dynamic Adaptive Aggregation Method for Visible and Infrared Image Target Detection and Recognition |
| WU Yuan,ZHU Peipei,LI Xue,XIE Xunwei,WANG Xianyuan |
| (National Key Laboratory of Complex Aviation System Simulation，Southwest China Institute of Electronic Technology,Chengdu 610036,China) |
| Abstract: |
| Multi-band image object detection and recognition is one of the important fundamental tasks,which aims to improve the target perception effect by supplementing and perfecting the target features through the imaging characteristics of different sensors,and has important practical significance for enhancing the intelligence level in fields such as transportation,medical,and military applications.Addressing the issue of poor detection and recognition results in object detection due to such factors as illumination,occlusion,and complex backgrounds,a dynamic adaptive aggregation visible and infrared image object detection and recognition method combining the complementary benefits of multi-source sensors is proposed.By designing a dynamic aggregation structure based on channel attention mixing and quadratic dynamic weight connection and using an improved path expansion Neck structure for semantic and spatial information interaction,this method fully explores the complementarity of multi-band images to improve the accuracy of multi-band image object fusion detection.The mean average precision(mAP) of the proposed method on three public datasets is more than 4 percentage points higher than that of the comparison model without the dynamic aggregation structure and the path expansion Neck. |
| Key words: multi-band image fusion recognition dynamic aggregation deep learning |
