| 引用本文: | 王睿,胡云雷,李海涛,高少泽,王刚.X射线焊缝图像缺陷实例分割算法[J].哈尔滨工业大学学报,2022,54(5):140.DOI:10.11918/202106010 |
| WANG Rui,HU Yunlei,LI Haitao,GAO Shaoze,WANG Gang.Defect segmentation algorithm for X-ray weld images[J].Journal of Harbin Institute of Technology,2022,54(5):140.DOI:10.11918/202106010 |
|
| 摘要: |
| 为了提高分割算法在焊缝缺陷检测工程中的实用性,提出轻量级焊缝缺陷评估网络MYNet。其中,轻量级的残差结构降低了模型的计算量,多层视觉融合机制FPN（feature pyramid network）提高了网络的分割能力,并行蒙版机制可以得到快速和高质量的缺陷分割蒙版；引入开源跨平台计算机视觉库OpenCV,利用像素阈值计算不同缺陷面积；引入腾讯超高性能的移动平台推理框架,加快模型在中央处理器的前向推理速度。搭建以ARM Cortex-A72架构为控制核心的数字化人工智能（artificial intelligence）评估设备,部署适用于缺陷检测的轻量级64位Linux系统,验证了焊缝缺陷评估算法的可行性。实验结果表明:本文模型能够有效定位和学习不同类型的缺陷特征；网络评估缺陷面积和位置信息的准确率为94.64%；相比于准确度较高但计算量较大的MS R-CNN网络,所提方法的准确率仅下降1.93%,但网络的参数权重仅为MS R-CNN网络的1/14,网络执行所需计算力更低。在基于ARM（advanced RISC machine）架构的低成本硬件上,轻量级的残差结构使网络前向推理速度提升了309%,仅用1.7 s完成低成本硬件上的焊缝实例分割任务。本文所提方法能有效学习和评估X射线焊缝缺陷图像,应用在评估设备上的算法降低了焊接质检的成本。 |
| 关键词: 焊接检测 缺陷分割 面积评估 智能设备 卷积神经网络 |
| DOI：10.11918/202106010 |
| 分类号:TG441 |
| 文献标识码:A |
| 基金项目:国家自然科学基金(62073118)； 河北省自然科学基金(F2019202305) |
|
| Defect segmentation algorithm for X-ray weld images |
|
WANG Rui1,2,HU Yunlei1,LI Haitao1,GAO Shaoze1,WANG Gang2
|
|
(1.School of Artificial Intelligence, Hebei University of Technology, Tianjin 300131, China; 2.State Key Laboratory of Advanced Welding and Joining (Harbin Institute of Technology), Harbin 150001, China)
|
| Abstract: |
| In order to improve the practicability of segmentation algorithm in weld defect detection, a lightweight weld defect evaluation network MYNet was proposed. In the network, the lightweight residual structure could reduce the amount of calculation of the model, the feature pyramid network (FPN) combined with multi-layer visual fusion mechanism could improve the segmentation ability of the network, and the parallel mask mechanism could obtain a fast and high-quality defect segmentation mask. The open source computer vision library platform OpenCV was introduced to calculate different defect areas by pixel threshold, and Tencent’s ultra-high-performance mobile platform reasoning framework was introduced to accelerate the forward reasoning speed of the model in the central processing unit. In this study, a digital artificial intelligence evaluation device was built with the ARM Cortex-A72 architecture as the control core, and a suitable lightweight 64-bit Linux system was deployed for defect detection to verify the feasibility of the proposed weld defect evaluation algorithm. Experimental results show that the model could effectively locate and learn different types of defect features. The network evaluated the defect area and location information with the accuracy of 94.64%. Compared with the MS R-CNN network which has high accuracy but requires a large amount of calculation, the accuracy of the proposed method was only reduced by 1.93%, while the parameter weight was only 1/14 of the MS R-CNN network, and the computing power required for network execution was lower. The lightweight residual structure increased the forward reasoning speed of the network by 309%, and it only took 1.7 s to complete the task of segmentation of weld instances on the low-cost hardware based on the advanced RISC machine (ARM) architecture. The method proposed in this paper can effectively learn and evaluate X-ray weld defect images, and the algorithm applied to the evaluation device reduced the cost of welding quality inspection. |
| Key words: welding detection defect segmentation area evaluation intelligent device convolutional neural network |
