电子测量与仪器学报

张吉祥,方东.基于YOLOv8-SPH的防震锤缺失检测[J].电子测量与仪器学报,2025,39(2):251-261

基于YOLOv8-SPH的防震锤缺失检测

Detection of missing shockproof hammers based on YOLOv8-SPH

DOI：

英文关键词:shockproof hammers deep learning multi-scale features YOLOv8n model pruning

基金项目:

作者	单位
张吉祥	三峡大学电气与新能源学院宜昌443002
方东	三峡大学机械与动力学院宜昌443002

Author	Institution
Zhang Jixiang	School of Electrical Engineering and New Energy,China Three Gorges University, Yichang 443002, China
Fang Dong	School of Mechanical and Power Engineering, China Three Gorges University, Yichang 443002, China

摘要点击次数: 20

全文下载次数: 63

中文摘要:

针对输电线路中防震锤尺寸小,图像背景复杂,防震锤缺失难以检测的问题,提出一种轻量化YOLOv8-SPH的防震锤缺失检测模型。通过在YOLOv8n网络的颈部引入160×160和320×320的浅层尺度特征图,并在检测头中融入相应尺度的目标检测模块,提升了特征图之间的上下文信息融合能力,有效扩大了模型的感受野,使得模型能够捕捉到更多防震锤缺失的特征语义信息。还创新性地提出了多尺度高效特征提取模块(MultFaster),通过部分卷积、多级特征提取和残差连接机制,在保持防震锤特征检测精度的同时,减少网络的计算量和参数量。此外,在颈部网络中引入动态上采样算子,提高重建特征图的分辨率,提高了该模型对防震锤缺失检测的精度,同时,将原模型解耦式检测头更换为轻量化检测头,降低了模型计算的复杂度并提升检测效率。最后对改进后的网络进行基于幅值的层自适应稀疏化剪枝,进一步减小模型参数及计算量。在针对自制防震锤缺失数据集的测试中,YOLOv8-SPH表现卓越,其mAP@0.5达到了91.51%,相比原始YOLOv8n提高了6.3%,参数量减少了80.73%,计算量减少了48.14%,模型尺寸减少了62.41%。该模型在计算量和参数量降低的同时,提高了检测精度,充分满足了对输电线路中的防震锤进行高效和准确检测的需求,具有实用性。

英文摘要:

To address challenges in detecting missing Shockproof Hammers on transmission lines due to their small size, complex image backgrounds, and subtle presence, this study proposes a lightweight YOLOv8-SPH model for damper absence detection. The model introduces shallow-scale feature maps of 160×160 and 320×320 within the neck of the YOLOv8n network and integrates multi-scale detection modules within the detection head. This enhances contextual information fusion across feature maps, effectively expanding the receptive field, enabling the model to capture richer semantic features related to damper absence. An innovative multi-scale high-efficiency feature extraction module (MultFaster) is also introduced, utilizing partial convolutions, multi-level feature extraction, and residual connections. This structure maintains detection accuracy for damper features while reducing computational complexity and parameter load. Additionally, a dynamic upsampling operator is incorporated into the neck network to improve feature map resolution, improving the model’s accuracy in detecting missing Shockproof Hammers. To further optimize, the original model’s decoupled detection head is replaced with a lightweight detection head, reducing computational complexity and boosting detection efficiency. The enhanced network undergoes amplitude-based layer-adaptive sparse pruning, significantly reducing model parameters and computational load. Testing on a custom damper absence dataset demonstrates YOLOv8-SPH exhibited remarkable performance, achieving an mAP@0.5 of 91.51%, which marks a 6.3% improvement over the original YOLOv8n. Additionally, parameter count is reduced by 80.73%, computational load by 48.14%, and model size by 62.41%. The model achieves improved detection accuracy while reducing computational complexity and parameter size, effectively meeting the demands for efficient and precise detection of Shockproof Hammers in transmission lines, showcasing significant practical value.

查看全文查看/发表评论下载PDF阅读器