Deep Learning-Based Crack Damage Detection Using Convolutional Neural Networks

提出一种无需人工设计特征即可检测混凝土裂缝的 CNN 分类器，在约 4 万张 256x256 图像块上训练，并通过滑动窗口应用于大幅照片。在一座未用于训练的结构的高分辨率照片上测试，包括眩光、阴影和发丝裂缝，其表现优于 Canny 和 Sobel 边缘检测。摘要中未提及无人机影像；这是一项地面相机研究，后来的无人机裂缝检测工作以此为基础。

主要发现

• About 98% accuracy on 256x256 patches (per abstract)
• Tested on 55 images of 5,888x3,584 px from a structure not used in training/validation
• Clearly better than Canny and Sobel edge detectors under glare, shadow and thin-crack conditions

局限

• Not UAV-specific per the abstract; images are not described as aerial
• Patch-level classification, not pixel-level crack width measurement
• Concrete cracks only; other damage types not covered