基于改进YOLOX-S的绝缘子图像识别技术研究

Research on insulator image recognition technology based on YOLOX-S

当电力巡检使用无人机对绝缘子进行航拍作业时,受各种因素影响,航拍图片中绝缘子尺寸小、分布密集,导致无法准确识别。为实现顺光、逆光状态下较小绝缘子的准确识别,提出一种基于YOLOX-S的绝缘子识别方法。首先,通过多角度航拍采集不同环境下的绝缘子图像进行预处理,获得训练样本;其次,用Mosaic-6进行数据增强处理,并传入网络进行训练,此方法计算Batch Normalization时一次采用6张图片的数据,mini-batch不需要很大,通过改进的YOLOX-S网络结构增加上采样和下采样次数进行特征融合,以提高网络对于小目标的特征提取能力;再次,将制作好的数据集输入网络训练;最后,在准确率、召回率和平均精度等方面与其他检测算法进行对比。结果表明:提出的改进算法对绝缘子识别的平均精度高达98.94%,与原YOLOX-S相比,提高了对较小绝缘子识别的准确率,从而有效提高了电力巡检绝缘子检测的准确率。

When drones are used for aerial photography of insulators in power inspection,affected by various factors,the insulators in the aerial pictures are small in size and densely distributed,which cannot be accurately identified.In order to realize the accurate identification of small insulators in the forward light and backlight,an insulator identification method based on YOLOX-S was proposed.Firstly,to obtain enough training samples,insulator images in different environments were collected through multi-angle aerial photography.And then,those images were processed by Mosaic-6 and trained in the network.In this way,mini-batch does not need to be large,because this method calculates the data of 6 pictures at the time when calculating Batch Normalization.In order to improve the feature extraction ability of the improved YOLOX-S network for small targets,feature fusion was done by increasing the number of up-sampling and down-sampling on the structure of this network.Then the prepared date was input into the network for training.Finally,the accuracy and recall rates were compared with another detection algorithms.The results show that the AP identified in this paper for insulators is 98.94%.The accuracy of identification of smaller insulators was improved when compared to original YOLOX-S.The improved insulator image recognition method of YOLOX-S proposed in this paper can effectively improve the accuracy of insulator detection for power inspection.