注意力机制和Faster RCNN相结合的绝缘子识别

PDF下载HTML

Title:

Insulator recognition based on attention mechanism and Faster RCNN

作者:

赵文清¹, 程幸福¹, 赵振兵², 翟永杰¹

1. 华北电力大学 控制与计算机工程学院, 河北 保定 071003;
2. 华北电力大学 电气与电子工程学院, 河北 保定 071003

Author(s):

ZHAO Wenqing¹, CHENG Xingfu¹, ZHAO Zhenbing², ZHAI Yongjie¹

1. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China;
2. School of Electrical and Electronic Engineering, North China Electric Power University, Baoding 071003, China

关键词:

Faster RCNN; 绝缘子; 注意力机制; SENet; 特征通道; RPN; 建议框; 特征向量

Keywords:

Faster RCNN; insulator; attention mechanism; SENet; characteristic channel; RPN; proposal boxes; feature vector

分类号:

TP391

DOI:

10.11992/tis.201907023

摘要:

针对利用Faster RCNN识别绝缘子图像过程中定位不够准确的问题，提出一种注意力机制和Faster RCNN相结合的绝缘子识别方法。在特征提取阶段引入基于注意力机制的挤压与激励网络（Squeeze-and-Excitation Networks，SENet）结构，使模型能够关注与目标相关的特征通道并弱化其他无关的特征通道；根据绝缘子的特点，对区域建议网络（region proposal network，RPN）生成锚点（anchor）的比例和尺度进行调整；在全连接层运用注意力机制对周围建议框的特征向量赋予不同权重并进行融合，更新目标建议框的特征向量。实验结果表明：与传统的Faster RCNN算法相比，改进后的算法能够较好地识别出绝缘子。

Abstract:

In order to solve the problem of inaccurate location in the process of recognizing insulator image using Faster RCNN, this paper proposes an insulator recognition method based on attention mechanism and Faster RCNN. Firstly, the Squeeze-and-Excitation Networks (SENet) structure based on attention mechanism is introduced in the feature extraction stage to enable the model to focus on the target-related feature channels and weaken other irrelevant feature channels. Then, according to the characteristics of insulators, the proportion and scale of anchors generated by regional proposal network (RPN) are adjusted. Finally, the attention mechanism is applied in the full connected layer to give different weights to the feature vectors of the surrounding suggestion boxes and fuse them to update the feature vectors of the target suggestion boxes. The experimental results show that the improved algorithm can recognize insulators better than the traditional Faster RCNN algorithm.