Autonomous detection of damage to multiple steel surfaces from 360° panoramas using deep neural networks

• Published in: Computer-aided civil and infrastructure engineering Vol. 36; no. 12; pp. 1585 - 1599
• Main Authors: Luo, Cai, Yu, Leijian, Yan, Jiaxing, Li, Zhongwei, Ren, Peng, Bai, Xiao, Yang, Erfu, Liu, Yonghong
• Format: Journal Article
• Language: English
• Published: Hoboken Wiley Subscription Services, Inc 01-12-2021
• Subjects: Artificial neural networks; Damage detection; Distortion; Inspection; Machine learning; Neural networks; Safety factors; Steel; Unmanned aerial vehicles
• ISSN: 1093-9687; 1467-8667
• DOI: 10.1111/mice.12686

Structural health assessments are essential for infrastructure. By using an autonomous panorama vision‐based inspection system, the limitations of the human cost and safety factors of previously time‐consuming tasks have been overcome. The main damage detection challenges to panorama images are (1) the lack of annotated panorama defect image data, (2) detection in high‐resolution images, and (3) the inherent distortion disturbance for panorama images. In this paper, a new PAnoramic surface damage DEtection Network (PADENet) is presented to solve the challenges by (a) using an unmanned aerial vehicle to capture panoramic images and a distorted panoramic augmentation method to expand the panoramic dataset, (b) employing the proposed multiple projection methods to process high‐resolution images, and (c) modifying the faster region‐based convolutional neural network and training via transfer learning on VGG‐16, which improves the precision for detecting multiple types of damage in distortion. The results show that the proposed method is optimal for surface damage detection.