基于改进YOLOv5的轻量化绝缘子表面缺陷检测

Detection of Surface Defects in Lightweight Insulators Using Improved YOLOv5

针对无人机检测缺陷绝缘子时,存在目标特征不明显、小目标检测效果差、无法同时满足检测速度和精度的问题,提出一种基于改进YOLOv5的绝缘子缺陷检测算法。首先,针对目标特征不明显的问题,将ConvNeXt网络应用到YOLOv5主干网络中,以加强网络特征提取能力;其次,针对图像中的小目标特征,在主干网络中引入坐标注意力机制,提高对小目标的检测精度;然后,对改进模型进行剪枝操作,剪去模型中冗余的通道,从而减少模型参数量,使模型更加轻量化。实验结果表明:所提算法在绝缘子缺陷数据集IDID上的平均精度均值达到93.84%,较原始算法提升了3.4个百分点;检测速率达到166 frame/s,较原算法速率提升了69.4%,可以满足对输电线路实时检测的要求。

Herein,an improved insulator defect-detection algorithm,YOLOv5,is proposed to overcome the shortcomings,including inconspicuous target features and poor detection of small targets when detecting trapped insulators using unmanned aerial vehicles,which cannot satisfy both detection speed and accuracy.First,ConvNeXt is applied to the YOLOv5 reference network to improve its ability to extract the features of obscure targets.Moreover,a coordinate attention mechanism is introduced into the reference network to improve its detection accuracy with respect to small targets in an image.Then,the improved model is pruned to eliminate its redundant channels,thus reducing the number of model parameters and making the model more lightweight.The experimental results show that the improved model achieves an average detection accuracy of 93.84%with respect to the insulator-defect dataset IDID,which is 3.4 percentage points higher than the accuracy achieved by the original algorithm.Moreover,the highest detection rate achieved by the proposed algorithm is 166 frame/s,which is 69.4%higher than that achieved by the original algorithm.These results prove that the improved algorithm meets the requirements of real-time transmission-line detection.