Crack is the most common defect in the railway sleeper inspection work. However, it is still lack of effective algorithms to automatically detect. Two deep learning based methods were popularly used to detect cracks: two-stage methods and one-stage methods. However, they both have their corresponding shortcomings: for the two-stage methods, they are too slow; for the one-stage methods, their accuracy is a problem. In this paper, we propose using a divide-and-conquer strategy of labels to improve the accuracy of the one-stage methods. A one-stage crack detector called CF-NET is proposed by us including two main innovations: a new detection pipeline (CF module) and modified loss function smooth-flat. Finally, the proposed model CF-NET achieves 98.1% accuracy with 17 FPS real-time speed. The accuracy of CF-NET is matched with the two-stage method Faster R-CNN, but faster at least 3\(\times \). The meaning of our work is that we provide a real-time and high-accuracy crack detector to better meet the actual demands of the railway sleeper crack detection task.

中文翻译：

---

使用一级检测器自动检测混凝土轨枕裂缝

裂缝是铁路轨枕检查工作中最常见的缺陷。但是，仍然缺乏有效的算法来自动检测。普遍使用两种基于深度学习的方法来检测裂纹：两阶段方法和一阶段方法。但是，它们都有各自的缺点：对于两阶段方法，它们太慢；对于一级方法，其准确性是一个问题。在本文中，我们提出使用标签的分而治之策略来提高一级方法的准确性。我们提出了一种称为CF-NET的单级裂纹检测器，其中包括两项主要的创新：新的检测管线（CF模块）和改进的损耗函数平稳平滑。最后，提出的CF-NET模型以17 FPS的实时速度达到98.1％的精度。\（\ times \）。我们的工作意思是我们提供一个实时，高精度的裂缝检测器，以更好地满足铁路轨枕裂缝检测任务的实际需求。