Multi-Type Bridge Damage Detection Method Based on YOLO

In this study, we try to solve the problem by using UAV and YOLO, which is one of the deep learning methods, for bridge inspection work, which is concerned about labor shortage and cost. By using deep learning, it is possible to shorten the work time by programming the work and suppress oversights and mistakes by add-ing an objective element to the diagnosis of damage. In the verification using CNN conducted by Tabata et al. As a previous study, the damaged part could be roughly recognized by the damage recognition of the image tak-en by the UAV, but the background part that had not been trained was mistakenly recognized as the damage. In addition, the UAV video diagnosis took a long time to detect, making it unsuitable for practical use. For these problems, we will verify using the YOLO v3 model, which is resistant to false detection of the background and can perform detection at high speed.

在本项研究中，我们尝试利用无人机（UAV）与YOLO（一种深度学习方法）相结合的方式，以解决桥梁检测工作中的劳动力短缺与成本问题。通过深度学习技术，我们得以通过编程优化工作流程，从而缩短作业时间，并通过引入客观诊断元素，有效抑制了检测过程中的疏漏与错误。借鉴Tabata等人的先前研究，通过CNN（卷积神经网络）进行的验证显示，无人机拍摄的图像损伤识别能够大致识别受损部分，然而，未经训练的背景部分却错误地被识别为受损区域。此外，无人机视频诊断所需时间过长，使其在实际应用中显得不甚适宜。针对上述问题，我们将采用对背景误检具有较强抵抗力的YOLO v3模型进行验证，该模型能够在高速检测的同时，避免背景误检。