| 基于非配对图像间转化弱监督学习的输电线路检测 |
投稿时间：2023-02-13 修订日期：2023-04-26  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 65 |
| 全文下载次数: 0 |
|
|
| 中文摘要:为了利用UAV航拍图像检测输电线路从而确保电力系统稳定运行,提出了一种基于弱监督学习和非配对图像间转化的输电线路检测方法。利用弱监督学习框架生成输电线路的定位掩码,通过引入新的并行扩展注意力(PDA)模块整合来自不同感受野的信息,从而重新校准通道重要性并提高检测精度。采用基于关联规则学习的算法生成伪线数据集,运用PDA中的注意力定位掩码(ALM)和伪线数据之间的非配对图像间转化技术构建精炼网络,从而增强输电线路的线形特性,实现了仅需图像级标签即可直接检测。实验结果表明,就F1分数而言,所提出的检测方法比目前最先进的递归噪声样本更新(RNLU)方法优越2.74%,并在消融实验中验证了精炼网络每个步骤都具有有效性。 |
| 中文关键词:弱监督学习 图像间转化 输电线路 无人机 注意力机制 |
| |
| Transmission Line Detection based on Weakly Supervised Learning of Unpaired Image Transformation |
|
|
| Abstract:In order to use UAV aerial images to detect transmission lines to ensure the stable operation of power system, a transmission line detection method based on weak supervised learning and non paired image transformation is proposed. The weak supervised learning framework is used to generate the location mask of the transmission line. By introducing a new parallel extended attention (PDA) module to integrate information from different receptive fields, the importance of the channel is recalibrated and the detection accuracy is improved. The algorithm based on association rule learning is used to generate pseudo wire data set, and the refining network is constructed by using the attention location mask (ALM) in PDA and the non paired image conversion technology between pseudo wire data, so as to enhance the linear characteristics of transmission lines and realize direct detection only by image level labels. The experimental results show that the proposed detection method is 2.74% better than the current most advanced recursive noise sample update (RNLU) method in terms of F1 score, and the ablation experiments verify that each step of the refining network is effective. |
| keywords:weak supervised learning Image to image conversion Transmission line UAV Attention mechanism |
| 查看全文 查看/发表评论 下载pdf阅读器 |
位访问者,今日访问
