[论文解读] Vision-based Structural Inspection using Multiscale Deep Convolutional Neural Networks

Vedhus Hoskere, Yasutaka Narazaki+2 more|arXiv (Cornell University)|May 2, 2018

Infrastructure Maintenance and Monitoring参考文献 26被引用 80

一句话总结

该论文提出一种像素级、多尺度的深度卷积神经网络方法，自动在图像中定位并分类六类结构损伤，输出分割的损伤区域。

ABSTRACT

Current methods of practice for inspection of civil infrastructure typically involve visual assessments conducted manually by trained inspectors. For post-earthquake structural inspections, the number of structures to be inspected often far exceeds the capability of the available inspectors. The labor intensive and time consuming natures of manual inspection have engendered research into development of algorithms for automated damage identification using computer vision techniques. In this paper, a novel damage localization and classification technique based on a state of the art computer vision algorithm is presented to address several key limitations of current computer vision techniques. The proposed algorithm carries out a pixel-wise classification of each image at multiple scales using a deep convolutional neural network and can recognize 6 different types of damage. The resulting output is a segmented image where the portion of the image representing damage is outlined and classified as one of the trained damage categories. The proposed method is evaluated in terms of pixel accuracy and the application of the method to real world images is shown.

研究动机与目标

• 推动在灾后 civil infrastructure 自动化、可扩展的检测，其中人工检测受限。
• 使用多尺度深度学习开发像素级损伤定位方法。
• 实现六种不同损伤类型的识别并生成带标签的分割损伤地图.

提出的方法

• 使用最先进的深度卷积神经网络在多个尺度执行像素级图像分类。
• 训练以识别六个损伤类别，并输出带有类别标签的分割图像。
• 提供损伤定位与分类流程，通过利用多尺度上下文比现有CV技术有改进。
• 通过真实世界图像的像素精度评估性能。

实验结果

研究问题

• RQ1 多尺度深度CNN是否能够在多种损伤类型中实现对结构损伤的像素级准确定位？
• RQ2 模型在后事件结构图像中对六个预定义损伤类别的分类与分割有多准确？
• RQ3 与单尺度方法相比，多尺度方法是否提高了损伤分割效果？

主要发现

• 该方法在多个尺度上实现对损伤的像素级分类。
• 输出是分割图像，其中损伤区域被轮廓化并标记为六个损伤类别之一。
• 该方法在真实世界图像上以像素准确度进行评估。