基于改进YOLOv5的电力金具目标检测方法

A Target Detection Method for Electric Power Hardware Based on Improved YOLOv5

为实现在复杂环境下对电力金具的准确快速检测,提出了一种基于改进YOLOv5(You Only Look Once version 5)的电力金具目标检测方法。针对现有目标检测方法参数量大、类别不平衡影响网络训练效果等问题,引入了轻量化网络Ghost Net,减少了冗余特征图对网络预测的影响,通过引入ECA(Efficient Channel Attention)注意力机制提高了网络的特征提取能力,避免了目标预测受到复杂环境的干扰,并使用Focal-CIoU(Focal-Complete-IoU) Loss损失函数提高了模型对少数类别的识别能力。实验结果表明,提出的算法在准确率和平均精度均值等评价指标中皆优于原始的YOLOv5算法,并且参数量仅占原始网络的61.2%。精确度提升了4.0%,mAP@0.5提升了1.9%,mAP@0.5:0.95提升了4.5%。同时,提出的算法在保持较高的预测精度的同时,均衡了推理速度,能够达到实时的电力金具目标检测要求。

To achieve accurate and fast detection of power fittings in complex environments,this paper proposes an improved YOLOv5(You Only Look Once version 5) based target detection method for power fittings.In view of the problem that the existing target detection methods have a large number of parameters and the imbalance of categories affect the network training effect,this paper introduces a lightweight network Ghost Net which reduces the impact of redundant feature mAPs on network prediction.The ECA(Efficient Channel Attention) attention mechanism is introduced to improve the network's feature extraction ability,avoid the interference of target prediction by complex environments,and use Focal-CIoU Loss function to improve the model's recognition ability for a few categories.The experimental results show that the algorithm proposed in this paper outperforms the original YOLOv5 algorithm in terms of accuracy and average accuracy,and its parameter count only accounts for 61.2% of the original network.Improved accuracy by 4.0%,mAP@0.5 An increase of 1.9%,mAP@0.5:0.95 increased by 4.5%.At the same time,the algorithm proposed in this article balances the inference speed while maintaining high prediction accuracy,and can meet the requirements of real-time power hardware target detection.