基于改进YOLOv5的遥感影像小目标识别研究

Research on Small Target Recognition in Remote Sensing Imagery Based on Improved YOLOv5

摘要: 由于小目标通常具有分辨率低、特征信息不足以及背景干扰严重等特点，传统目标检测方法在该任务上的性能往往受限。针对上述问题，本文提出了一种基于改进YOLOv5的遥感影像小目标智能检测方法，旨在提升检测精度与召回率。本文的主要改进体现在三个方面。首先，设计了空间金字塔空洞卷积模块(SPD-Conv)，通过多尺度空洞卷积有效扩大感受野，同时保持特征图的高分辨率，从而减少下采样过程中小目标细节信息的丢失。其次，引入轻量级高效通道注意力机制(ECA-Net)，以自适应方式增强小目标相关特征表达，并抑制背景噪声干扰。最后，采用双向特征金字塔网络(BiFPN)替代传统路径聚合网络，实现多尺度特征的双向高效融合，显著提升深层特征的语义表达能力。在公开遥感数据集AI-TOD上的实验结果表明，所提方法在小目标检测任务中表现优异，平均精度均值(mAP)达到36.8%，较基线模型YOLOv5提升了7.3个百分点。实验结果充分验证了所提方法在复杂遥感场景中进行小目标检测的有效性与优越性。

Abstract: Small object detection has always been the core challenge in computer vision in remote sensing images. Due to the characteristics of small object detection, such as low resolution, scarcity of feature information and serious background interference, the traditional object detection algorithm performs poorly in this task. This paper proposes an intelligent detection system for remote sensing images of small objects based on improved YOLOv5, which aims to improve the detection accuracy and recall rate of small targets. The main innovations of this study include three aspects: First, a spatial pyramid depth convolution module (SPD-Conv) is designed, which effectively expands the sensing field through a multi-scale hollow convolution structure, while maintaining the high resolution of the feature map to avoid losing detailed information of small objects during the downsampling process. Secondly, a lightweight channel attention mechanism (ECA-Net) is introduced to adaptively enhance the expression of characteristics related to small objects and suppress interference from irrelevant background information. Third, the two-way feature pyramid network (BiFPN) is used to replace the traditional path aggregation network, which realizes the two-way integration of multi-scale features and greatly improves the semantic representation ability of small objects in deep networks. Experimental results on the public remote sensing data set AI-TOD show that this method performs well in small object detection tasks, with a detection accuracy rate (mAP) of 36.8%, 7.3 percentage points higher than the original YOLOv5, verifying the effectiveness of the proposed improvement strategy.