基于改进YOLOv3的变压器定位检测研究

Transformer Detection Based on Improved YOLOv3

目前目标检测算法由于数量和背景不平衡导致的电力部件检测精度不高,同时在背景复杂的情况下分类定位不准确。针对上述问题,提出一种改进的YOLOv3的变压器检测算法,使用DarkNet-53深度卷积神经网络作为抽取图像特征的骨干网络。使用Focal loss和均衡交叉熵函数改进原YOLOv3的损失函数,使模型在样本数量少的类别上倾注更多的注意力;使用K-means对样本中的变压器框进行聚类生成预选框,并调整YOLOv3输出结构使其适应变压器目标检测任务,提高检测的效率。实验结果表明改进后的YOLOv3在变压器检测上有更高的准确率和速率。

At present,the detection accuracy of power components due to the imbalance of the number and background of the target detection algorithm is not high,and the classification and positioning are not accurate in the case of complex background.In view of the above problems,an improved YOLOv3 transformer detection algorithm is proposed,using DarkNet-53 deep convolutional neural network as the backbone network to extract image features.Focal loss and balanced cross-entropy function is used to improve the original YOLOv3 loss function,so that the model can pay more attention to the category with a small number of samples;K-means is used to cluster the transformer boxes in the samples and generate pre-selection boxes,and the output structure of YOLOv3 is adjusted to make it suitable for the target detection task of transformers and improve the detection efficiency.The experimental results show that the improved YOLOv3 has higher accuracy and rate in transformer detection.