An Efficient Augmented Reality (AR) System for Enhanced Visual Inspection
An Efficient Augmented Reality (AR) System for Enhanced Visual Inspection
复制标题
用于增强视觉检查的高效增强现实 (AR) 系统
DOI：
10.12783/shm2019/32278
复制
发表时间：
2019
影响因子：
6.6
通讯作者：
F. Yuan
中科院分区：
文献类型：
--
作者：
Shaohan Wang;Sakib Ashraf Zargar;C. Xu;F. Yuan
虽然结构的人工视觉检查具有相对简单和低成本的优点，但它通常是耗时的、劳动密集型的和高度主观的。增强现实（AR）由于其能够实时地向用户提供关于工作环境的附加信息，因此在过去已经被用于通过在检查过程期间支持人类工作者来解决人工视觉检查的一些限制。本文介绍了一种高效的基于深度学习（DL）的增强现实（AR）系统的开发，用于识别与结构原始状态的关键偏离，重点关注两种异常类别-腐蚀和疲劳裂纹。大多数常见的AR设备通常带有用于捕获图像/视频数据的内置摄像头，存储器和微处理器。然而，由于处理能力有限，底层深度学习（DL）模型必须首先在外部进行训练，然后在设备上本地部署合适版本的训练模型。然后，模型输出用于识别与结构的原始状态的临界偏离的信息，突出腐蚀区域、疲劳裂纹和/或两者的组合。该信息通过头戴式或手持AR设备在当前视场上真实的时间叠加，以便增强人类视觉。然后，工人可以专注于突出显示的区域进行更详细的检查。所提出的AR系统的可行性证明了使用实验室检查常见的机械部件，如管道，板等，为了使模型保持学习的基础上的AR眼镜的输入，联邦学习的策略介绍了对文件的结尾。
While manual visual inspection of structures has the advantage of being relatively simple and low cost, it is usually time consuming, labor intensive and highly subjective. Augmented reality (AR), because of its ability to provide the user with additional information about the working environment in real-time, has been used in the past to address some of the limitations of manual visual inspection by supporting human workers during the inspection process. The paper presents the development of an efficient deep learning (DL) based augmented reality (AR) system for identifying critical departures from the pristine state of the structure with focus on two anomaly categories- corrosion and fatigue cracks. Most of the common AR devices usually come with a built-in camera for capturing image/video data, a storage and a microprocessor. However, due to the limited processing power, the underlying deep learning (DL) model has to be first trained externally and a suitable version of the trained model is then deployed locally on the device. The model then outputs information for identifying critical departures from the pristine state of the structure e.g., highlighting corroded regions, fatigue cracks and/or combination of both. This information is overlaid real time over the current field of view through either a headmounted or a hand-held AR device in order to augment the human vision. The worker can then focus on the highlighted region for a more detailed inspection. The feasibility of the proposed AR system is demonstrated using laboratory inspection of common mechanical components likes pipes, plates etc. In order to enable the model to keep learning based on the inputs from the AR glasses, a strategy for federated learning is introduced towards the end of the paper.