基于多重注意力和特征对齐的销钉缺陷检测方法

Bolt Defect Detection Method Based on Multiple Attention and Feature Alignment

电网规模的快速扩张对输电线路的检测维护提出更高的要求,智能高效的自动化电力巡检图像识别算法作为提高检测维护的手段,已经成为目前研究的重要方向之一。为实现巡检图像中销钉缺陷目标的准确识别,文章提出一种基于多重注意力和特征对齐的销钉缺陷检测方法。首先,构建基于可变形注意力的可变形检测变压器(deformable detection transformer,Deformable-DETR)框架,解决现有识别算法中无法通过卷积神经网络进行像素点轮廓-环境关系建模的问题;之后,提出区域注意力特征提取模块,解决由可变形注意力带来的特征粒度不足的问题;为丰富目标物体特征质量,提出基于注意力的候选框特征对齐模块和约束函数;最后,选取华中地区某电力公司的航拍图像数据验证算法性能。结果表明,所提算法总体性能相比于现有的卷积算法提升5.4%,总体平均检测精度达到90.5%。

The rapid expansion of power grid had higher requirements for the detection and maintenance of transmission lines,thus intelligent and efficient automated power inspection algorithms become one of the important research directions.In order to accurately identify the defective bolts in power line inspection image,we proposed a bolt defect detection method based on multiple attention and feature alignment.Firstly,we built a Deformable-DETR(deformable detection transformer)framework based on deformable attention to solve the problem that existing power inspection algorithms could not model the pixel contour-environment relationship through convolutional neural networks.Secondly,we proposed a proposal based local attention module to alleviate the problem of insufficient feature granularity caused by deformable attention.Thirdly,in order to enrich the object feature quality under the existing data,we proposed an object-level based feature alignment module based and a constraint functions.Finally,drone inspection image of a power company in central China were selected for verification.The experimental results show that the overall performance of the proposed algorithm improves by 5.4%compared with the existing convolutional algorithms,and the overall mean average precision reaches 90.5%.