改进YOLOv10n的复杂场景设备检测算法

Improved YOLOv10n-Based Device Detection Algorithm for Complex Scenes

摘要: 针对加速器中子源设备巡检场景中结构复杂、设备类型多样且易受遮挡干扰等问题，现有目标检测方法在检测精度与稳定性方面仍存在不足。为此，本文提出了一种基于改进YOLOv10的复杂场景设备检测算法：ECS-YOLO。首先，在特征提取阶段引入EffectiveSE注意力机制，对通道特征进行自适应重标定，增强关键设备特征表达能力。其次，设计C2f_RFAConv模块，将RFA注意力机制嵌入C2f结构，在不显著增加计算开销的前提下提升多尺度特征提取与感知能力。最后，在损失函数中引入SIoU损失，从距离、角度与形状匹配等多维度约束边界框回归过程，提高目标定位精度。实验结果表明，与基线模型相比，改进的ECS-YOLO模型在mAP@50、P、R上分别提高了3.7%，3.5%，4.8%。

Abstract: In accelerator neutron source equipment inspection scenarios, the complex structural environment, diverse equipment types, and frequent occlusion pose significant challenges to existing object detection methods, leading to insufficient detection accuracy and stability. To address these issues, this paper proposes an improved YOLOv10-based device detection algorithm for complex scenes, termed ECS-YOLO. First, the EffectiveSE attention mechanism is introduced in the feature extraction stage to adaptively recalibrate channel features, thereby enhancing the representation of critical device features. Second, a C2f_RFAConv module is designed by embedding the RFA attention mechanism into the C2f structure, which improves multi-scale feature extraction and perception capability without significantly increasing computational overhead. Finally, the SIoU loss function is incorporated to constrain bounding box regression from multiple aspects, including distance, angle, and shape matching, further improving localization accuracy. Experimental results demonstrate that, compared with the baseline model, ECS-YOLO achieves improvements of 3.7%, 3.5%, and 4.8% in mAP@50, Precision, and Recall, respectively.