基于改进YOLOv13的小目标检测研究

Research on Small Object Detection Based on Improved YOLOv13

摘要: 针对图像中小目标分布密集、易发生重叠且受背景干扰，以及类别不平衡导致检测性能下降等问题，文章提出了一种基于改进YOLOv13的目标检测算法。首先，在骨干网络和颈部网络中，引入上下文引导块CGblock模块并替换所有C3K2模块，形成CG-C3K2模块，使检测更聚焦小目标区域，自适应地增强关键信息的表达能力，克制感受野受限问题，加强对小目标的特征增强；其次，在颈部结构中融合通道–空间协同注意力机制(Channel-Spatial Collaborative Attention Mechanism, CSCAM)，使特征融合过程更加聚焦于有效目标区域，提升对小目标和密集目标的检测效果；最后，采用改进的Wise-PIoU复合损失函数替代原YOLOv13模型中的损失设计，使得小目标能够更好地回归到真实标注框，有效提高了训练稳定性与检测精度。基于VOC2007/2012数据集的实验结果表明，所提出的改进算法较原始YOLOv13模型的mAP@0.5提升了6.6%，精确率提升了4.7%，召回率提升了5.8%。研究结果验证了改进的YOLOv13算法在复杂场景下对小目标的检测性能得到了有效提升。

Abstract: To address the challenges of dense small object distribution, severe occlusion, background interference, and category imbalance that degrade detection performance, this paper proposes an improved YOLOv13-based object detection algorithm. First, the Contextual Guided Block (CGblock) is introduced into both the backbone and neck networks, replacing all C3K2 modules to form CG-C3K2 modules, which adaptively enhance the representation of critical small-object regions, mitigate receptive field limitations, and strengthen feature expression for small targets. Second, a Channel-Spatial Collaborative Attention Mechanism (CSCAM) is integrated into the neck structure to focus feature fusion on effective target areas, thereby improving detection accuracy for small and densely packed objects. Finally, an improved Wise-PIoU composite loss function replaces the original loss design in YOLOv13, enabling more accurate regression of small objects to ground truth boxes and enhancing both training stability and detection precision. Experimental results on the VOC2007/VOC2012 datasets demonstrate that the proposed algorithm achieves a 6.6% increase in mAP@0.5, along with 4.7% and 5.8% improvements in precision and recall, respectively, compared to the original YOLOv13 model. The results validate the effectiveness of the improved algorithm in enhancing small-object detection performance in complex scenarios.