【目的】针对绝缘子缺陷尺寸小、检测时易受复杂背景干扰以及基线模型参数量大等问题,本文提出一种基于CPLC-YOLOv8改进的轻量化绝缘子缺陷检测算法。【方法】首先,设计轻量化RepNCSPELAN4-CAA替换YOLOv8主干网络中的C2f模块,降低参数量并增强特征表达能力;其次,新增小缺陷检测层P2,强化浅层与深层特征的融合,减少小目标信息的流失;然后,设计一种轻量化检测头,采用1×1卷积调整通道维度,并利用细节增强卷积替代传统3×3卷积,实现参数共享与特征增强;最后,引入卷积注意力机制,通过通道与空间双重注意力机制抑制背景干扰,增强关键特征表达,提升模型鲁棒性与检测精度。【结果】在自建绝缘子缺陷数据集上的试验结果表明,CPLC-YOLOv8的mAP@0.5达到0.928,相较于YOLOv8提升2个百分点;其模型参数量仅为1.72 MB,较YOLOv8减少42.8%;模型大小为4.12 MB,压缩31.3%。在多种经典网络模型对比中,CPLC-YOLOv8在检测精度、参数量和模型体积方面均表现出显著优势,尤其在小目标检测任务中展现出更强的鲁棒性和泛化能力。【结论】本文所提算法在保持高检测精度的同时,实现了模型的轻量化设计,适用于资源受限的边缘设备部署,具有良好的工程应用前景。未来工作将进一步探索多尺度特征融合与轻量化技术的结合,持续提升算法在实际电力巡检场景中的适应性与稳定性。

[Objective] Aiming at the problems of small insulator defect size, susceptibility to complex background interference during detection, and large parameter volume of the baseline model, this paper proposes a lightweight insulator defect detection algorithm based on an improved CPLC-YOLOv8. [Methods] Firstly, the lightweight RepNCSPELAN4-CAA module was designed to replace the C2f module in YOLOv8’s backbone network, reducing parameter quantity while enhancing feature representation capability. Secondly, a small-defect detection layer P2 was added to strengthen the fusion of shallow and deep features, minimizing the loss of small-target information. Subsequently, a lightweight detection head was developed, where 1×1 convolution was employed for channel dimension adjustment and detail-enhanced convolution was utilized to replace conventional 3×3 convolution, achieving parameter sharing and feature enhancement. Finally, the convolutional block attention mechanism was introduced to suppress background interference through dual channel-spatial attention mechanisms, enhancing key feature representation and improving model robustness and detection accuracy. [Results] Experimental results on the custom insulator defect dataset demonstrated that the proposed CPLC-YOLOv8 achieved a mAP@0.5 of 0.928, representing a 2 percentage point improvement over the original YOLOv8. The model parameters were reduced to only 1.72 MB (42.8% reduction compared to YOLOv8), with a compressed model size of 4.12 MB (31.3% compression). Comparative evaluations with classic network models confirmed that CPLC-YOLOv8 exhibited significant advantages in detection accuracy, parameter efficiency, and model compactness, particularly demonstrating superior robustness and generalization capability in small object detection tasks. [Conclusion] The proposed algorithm achieves lightweight model design while maintaining high detection accuracy, making it suitable for deployment on resource-constrained edge devices with promising engineering application prospects. Future work will further explore the integration of multi-scale feature fusion and lightweight techniques to continuously enhance the algorithm’s adaptability and stability in practical power inspection scenarios.

辽宁省教育厅项目(LJKFZ20220190)