基于多光谱成像的钢轨表面缺陷在线检测方法

The surface defect of rails is an important factor affecting the safety of railway operation. It is essential to inspect the surface defects strictly in the rail production. The existing manual way of surface inspection of rails is inefficient, needs high labor intensity and prone to missing detection of defects. Thus, manual inspection has become a "bottleneck" restricting rail production. Due to the complexity of the rail shape and large number of scales on the rail surface, the traditional two-dimensional detection method will lead to missing detection and a lot of false positives. Two-dimensional detection and three-dimensional detection methods have complementarity in surface detection, as they have own advantages and disadvantages. Combination of two-dimensional detection and three-dimensional detection can improve the accuracy of defect detection, but there is no feasible solution. Multispectral imaging technology is introduced in the project. Red, green, blue and near infrared channel images are separated from multispectral images captured by a multispectral camera. 3D surface data are reconstructed from red, green and blue channel images, and near infrared channel images are used as the gray-level images. The depth information and gray-level information are obtained synchronously with one camera. Open and closed surface defects of rails are detected by fusion of depth information and gray-level information, and reliable surface inspection of rails are realized. This project is of great significance for improving the production capacity of rails and ensuring safe operation of railways, and will bring great economic and social benefits.

钢轨表面缺陷是影响铁路运营安全的重要因素，在钢轨生产中必须对表面缺陷进行严格的检测。现有的人工检测钢轨表面缺陷方式效率低，工人劳动强度大并容易出现漏检，已经成为制约钢轨生产的“瓶颈”。由于钢轨端面形状的复杂性，并且表面覆盖大量氧化铁皮，传统的两维图像检测方式会造成缺陷的漏检与大量误报。两维检测与三维检测方法在表面检测上有各自的优势与不足，两者具有互补性，将两种方法相结合可提高缺陷检测的准确率，但是缺乏可行方案。本项目将采用多光谱成像技术，从多光谱相机采集的图像中分离出红、绿、蓝、近红外通道图像，通过红、绿、蓝通道图像重构表面三维数据，近红外通道图像作为表面灰度图像，由同一台相机同步获取表面深度与灰度信息，并通过深度与灰度信息融合方法检测钢轨表面开放式缺陷与封闭式缺陷，实现钢轨表面缺陷可靠检测。本项目对于提高钢轨的生产能力和保障铁路的安全运营具有重要的意义，从而创造巨大的经济效益和社会效益。

钢轨表面缺陷是影响铁路运营安全的重要因素，在钢轨生产中必须对表面缺陷进行严格的检测。现有的人工检测钢轨表面缺陷方式效率低，工人劳动强度大并容易出现漏检，已经成为制约钢轨生产的“瓶颈”。由于钢轨端面形状的复杂性，并且表面覆盖大量氧化铁皮，传统的两维图像检测方式会造成缺陷的漏检与大量误报。本项目采用多光谱成像技术，从多光谱相机采集的图像中分离出红、绿、蓝、近红外通道图像，利用红、绿、蓝通道图像重构表面三维数据，并通过二维信息与三维信息的融合检测钢轨表面缺陷，以提高钢轨表面缺陷检测的准确率与可靠性。本项目主要研究内容如下：1）研究了基于光度立体的钢轨表面缺陷在线三维检测方法，提出了一种能同时获取表面缺陷二维和三维信息的方法；2）构建了基于深度学习的钢轨表面二维缺陷检测模型，用于钢轨表面缺陷的在线检测；3）开发了一套二维与三维相融合的钢轨表面缺陷检测系统，应用于邯钢钢轨生产线。本项目取得的创新性成果如下：1）提出了一种适用于钢轨异形表面的光度立体表面三维重建方法，可在线获取钢轨表面缺陷的深度信息；2）提出了一种基于深度学习的钢轨表面缺陷识别方法，并采用半监督学习方法解决了钢轨缺陷样本数据量不足的问题；3）提出了一种基于全卷积神经网络（FCN）的二维与三维信息融合方法，提高了表面缺陷识别准确率与可靠性。本项目已开发了一套钢轨表面缺陷在线检测系统，并应用于邯钢钢轨生产线，经现场使用验证，可检测钢轨表面轧疤、压痕、凹坑等开放式缺陷和折叠、裂纹、夹杂等封闭式缺陷，缺陷检出率达99%以上，识别率达90%以上；满足生产线最高速度150m/min的检测要求。本项目对于控制钢轨表面质量，保障钢轨服役性能和高铁运行安全具有重大意义。本项目成果还可应用于棒线材、钢管、H型钢等长材生产线，对于提高这些生产线上产品质量、减少由于表面缺陷造成的损失起到重要作用。