基于深度学习的航天密封圈表面缺陷检测

针对航天密封圈表面缺陷人工检测效率低、传统图像处理检测算法通用性差的问题,提出了两种基于深度学习的密封圈表面缺陷检测算法。首先,针对缺陷大部分为小目标的特点,选取对小目标较敏感的RetinaNet网络作为检测算法的基本架构,通过在RetinaNet网络中引入轻量级网络MoGaA构建出MoGaA-RetinaNet算法。然后,为了提高检测精度,在MoGaA-RetinaNet基础上,用分解卷积模块代替MoGaA骨干网络中的深度卷积构建了newMoGaA骨干网络,设计出newMoGaA-RetinaNet算法。最后,在测试集上的实验结果表明,MoGaA-RetinaNet算法比RetinaNet算法检测速度更快,但检测准确率略低;而newMoGaA-RetinaNet算法实现了检测精度与检测速度的良好平衡,比RetinaNet算法准确率提升4.5%,达到92%,检测速度提升55%,达到31 frame/s,网络参数量减少50%。所设计的newMoGaA-RetinaNet算法可以实现密封圈表面缺陷的快速准确检测。

Surface defect detection of aerospace sealing rings based on deep learning

Aiming at the problems of low aerospace seal ring surface defect detection efficiency of manual inspection and poor versatility of traditional image processing detection algorithms, two kinds of deep learning based surface defect detection algorithms for aerospace sealing rings are proposed. Firstly, aiming at the characteristic that most of the defects are small targets, the RetinaNet network that is more sensitive to small targets is selected as the basic architecture of the defect detection algorithm, and the MoGaA-RetinaNet algorithm is constructed by introducing the lightweight network MoGaA into the RetinaNet network. Secondly, in order to improve the detection accuracy, on the basis of MoGaA-RetinaNet, the newMoGaA backbone network is constructed using the decomposition convolution module to replace the depthwise convolution in the MoGaA backbone network, and the newMoGaA-RetinaNet algorithm is designed. Finally, the experiment results on the test set show that the MoGaA-RetinaNet algorithm has faster detection speed but slightly lower detection accuracy compared with the RetinaNet algorithm;the newMoGaA-RetinaNet algorithm achieves a good balance of detection accuracy and detection speed, Compared with those of RetinaNet algorithm, the detection accuracy rate increases by 4.5%, reaches to 92%;the detection speed increases by 55%, reaches to 31 frame/s;and the number of network parameters is reduced by 50%. The newly designed newMoGaA-RetinaNet algorithm can achieve fast and accurate detection of the seal ring surface defects.