基于改进U-Net的金具图像小样本识别算法研究
Research on few-shot recognition algorithm of fittings image based on improved U-Net
谢智慧 1王文爽 2刘雪峰1
作者信息
- 1. 青岛科技大学自动化与电子工程学院 青岛 266061
- 2. 山东省电力公司烟台供电公司互联网部 烟台 264000
- 折叠
摘要
电力金具巡检是保证电网安全运行的关键任务.针对因金具样本类别不平衡、金具图像背景复杂而导致的误检、漏检问题,提出了一种改进U型网络(U-shaped network,U-Net)的检测方法.首先,通过生成对抗网络生成虚拟金具样本扩充数据集,解决数据集中样本类别不平衡的问题;然后,提出一种前景增强方法,在网络输出的特征图中加入背景掩膜,并优化损失函数;最后,将注意力机制嵌入U-Net,以提高模型在复杂背景下提取金具特征的能力.经实验证明,改进算法对电力金具目标的检测效果良好,其金具检测准确率达到98.82%,平均交并比达到83.94%,精确率达到91.01%,召回率达到86.18%,平均精度均值达到89.73%.改进算法不仅可应用于正常金具的检测,还有效适用于生锈金具的检测,为电力金具智能化检测提供了一种新思路.
Abstract
Power fittings inspection is a critical task in ensuring the safe operation of the power grid.To address the challenges of imbalanced fittings samples and complex background images leading to false and missed detections,an improved detection method based on the U-Net is presented.Firstly,a generative adversarial network is employed to generate synthetic fittings samples,alleviating the issue of imbalanced sample distribution in the dataset.Secondly,a foreground enhancement method is proposed,which applies a background mask to the feature map generated by the network and optimizes the corresponding loss function.Finally,an attention mechanism is integrated into the U-Net network to enhance the model's ability to extract fittings features in complex backgrounds.Experimental results demonstrate the effectiveness of the proposed algorithm in detecting fittings objects,the fittings detection accuracy reached 98.82%,the mean intersection over union reached 83.94%,the precision reached 91.01%,the recall reached 86.18%,and the mean average precision reached 89.73%.The proposed algorithm is not only applicable to normal fittings,but also effective in detecting rusty fittings.This approach provides a new perspective for the intelligent detection of fittings.
关键词
电力金具/U-Net/生成对抗网络/前景增强/ACmix/智能巡检Key words
fittings/U-Net/generative adversarial network/foreground enhancement/ACmix/intelligent inspection引用本文复制引用
基金项目
国家自然科学基金(61971253)
山东省自然科学基金(ZR2020MF011)
出版年
2024
NETL
NSTL
万方数据