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

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