Across a range of manufacturing contexts, automated quality control has been gaining significant attention because it offers competitive advantages such as cost reduction, high accuracy in defect detection, and system stability over time. Although computer vision has historically been the most commonly applied method in this context, novel approaches such as deep learning have recently become more frequent and are used in cases where traditional methods cannot be applied. Because of the surface texture and curvature of many metallic parts, detection of defects such as scratches, cracks, and dents can be challenging for traditional computer vision methods. In this study, an image acquisition system supported by a special lighting device that provides processable images from an extremely reflective cylindrical metallic surface has been developed. Multiple images obtained from a single lateral line of the surface, which is rotated at a specified speed, are combined using photometric stereo and given as input to a convolutional neural network that is employed to classify defective and non-defective samples. The results obtained from this method are close to 98.5% accurate.

中文翻译：

---

使用深度学习的在线金属表面缺陷检测

在整个制造环境中，自动化质量控制已获得了广泛的关注，因为它具有竞争优势，例如降低成本，缺陷检测的高精度和随时间推移的系统稳定性。尽管从历史上讲，计算机视觉一直是这种情况下最常用的方法，但是诸如深度学习之类的新方法近来变得越来越普遍，并在无法应用传统方法的情况下使用。由于许多金属零件的表面纹理和曲率，对于诸如划痕，裂缝和凹痕之类的缺陷的检测对于传统的计算机视觉方法可能具有挑战性。在这项研究中，开发了一种由特殊照明设备支持的图像采集系统，该系统可从极度反射的圆柱形金属表面提供可处理的图像。从表面的单条横向线获得的多个图像以指定的速度旋转，并使用光度立体图像进行组合，并作为卷积神经网络的输入，该卷积神经网络用于对有缺陷和无缺陷的样本进行分类。通过这种方法获得的结果准确度接近98.5％。