基于深度神经网络的航空叶片表面缺陷检测算法  

Aircraft blade surface defect detection based on deep neural networks

作　　者：苏宝华 张吟龙 张男 冯选 SU Baohua;ZHANG Yinlong;ZHANG Nan;FENG Xuan(Inspection and Testing Centering,Shenyang Liming Aero-Engine(Group),Corporation LTD.,Shenyang,Liaoning110043,China;Department of Network and Control System,Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang,Liaoning 110169,China;College of Information Engineering,Shenyang Universityof Chemical Technology,Shenyang,Liaoning 110142,China)

机构地区：[1]中国航发沈阳黎明航空发动机有限责任公司产品检验检测中心,辽宁沈阳110043 [2]中国科学院沈阳自动化研究所工业控制网络与系统研究室,辽宁沈阳110169 [3]沈阳化工大学信息工程学院,辽宁沈阳110142

出　　处：《光电子.激光》2025年第2期130-135,共6页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(62273332);中国航发黎明科研项目(KT23Y031-YIY)资助项目。

摘　　要：航空叶片表面缺陷的准确检测是保证航空发动机安全可靠运转的关键。目前基于视觉的航空叶片表面缺陷检测算法存在实时性差、漏检率高、定位目标不准等问题。针对上述问题,本文提出一种基于深度神经网络的航空叶片表面缺陷检测算法。为了提高检测实时性,本文设计了深度可分离卷积(depthwise separable convolution,DSC)模型分解标准卷积;为了降低小目标缺陷漏检,本文设计了SE-PAN(squeeze-and-excitation path aggregation network)模型对每个通道的特征进行重标定,使得具有更强信息的特征得到更多的关注;为了提高定位准确度,本文设计了Focal-DIOU(focal-distance intersection over union)损失函数减弱低效框的作用。在本文的航空叶片表面缺陷数据集上的实验证明:本文算法的Precision、Recall、AP达到了95.7%、94.6%、96.3%,检测帧率达到24帧/s,均优于主流检测算法。Accurate detection of surface defects on aircraft blades is crucial for ensuring the safe and reliable operation of aero-engines.Currently,vision-based algorithms for detecting surface defects on aircraft blades suffer from poor real-time performance,high missed detection rates,and inaccurate target localization.To address these issues,this paper proposes an aircraft blade surface defect detection algorithm based on deep neural networks.To improve detection real-time performance,we design the depthwise separable convolution(DSC)model to decompose standard convolutions.To reduce missed detection of small defect targets,we propose the squeeze-and-excitation path aggregation network(SEPAN)model to recalibrate the features of each channel,allowing features with stronger information to receive more attention.To enhance localization accuracy,we design the focal-distance intersection over union(Focal-DIOU)loss function to mitigate the effect of inefficient boxes.Experimental results on our aircraft blade surface defect dataset demonstrate that our algorithm achieves Precision,Recall and APof 95.7%,94.6%and 96.3%,respectively,with a detection frame rate of 24frames per second,all of which outperform mainstream detection algorithms.

关 键 词：缺陷检测 深度神经网络 损失函数 注意力模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]