Domain-Shift-Dataset of Defects on Metallic Surfaces (MSD-Shift)
收藏资源简介:
The dataset maps two different surfaces (domains) from mechanical engineering (Surfaces of Ball Screw Drives (BSD); Surface of Metallic Semi-finished Products (SEV)). The domains each contain image data with and image data without surface defects. The surfaces differ, but the defect features are similar across the domains. The dataset is thus suitable for investigating questions in the context of domain shift, domain generalization, and transfer learning. The dataset is structured as follows: BSD with defect (5240); BSD without defect (1896) SEV with defect (2018); SEV without defect (21806). Attention: The Domains are not balanced over classes. The SEV images are each 224x224 pixel RGB PNG files. THE BSD images are each 150x150 pixel RGB PNG files. The SEV data are excerpts from the Severstal dataset.
本数据集涵盖机械工程领域的两类不同表面(域):滚珠丝杠副表面(Ball Screw Drives,BSD)与金属半成品表面(Metallic Semi-finished Products,SEV)。两类域各自均包含带有表面缺陷和无表面缺陷的图像数据。两类表面本身存在差异,但不同域之间的缺陷特征具有相似性。因此该数据集适用于开展域偏移(domain shift)、域泛化(domain generalization)以及迁移学习相关问题的研究。 数据集的结构如下:带缺陷的BSD样本(5240张)、无缺陷的BSD样本(1896张)、带缺陷的SEV样本(2018张)以及无缺陷的SEV样本(21806张)。请注意:两类域在类别分布上并不均衡。SEV图像均为224×224像素的RGB格式PNG文件,BSD图像均为150×150像素的RGB格式PNG文件。SEV数据集的图像为Severstal数据集的节选样本。




