"Semi-AD (Semiconductor Anomaly Detection) dataset"

"The Semi-AD (Semiconductor Anomaly Detection) dataset is a specialized benchmark for unsupervised comparison-based anomaly detection (AD) in the semiconductor manufacturing. This dataset addresses the critical challenge of identifying rare defects in intricate circuit patterns by leveraging the inherent repetitive structures of semiconductor devices. It provides precisely aligned image pairs consisting of a reference device and a test device and ground truth annotations of defective regions are derived from pixel-wise differences between the aligned pairs.The dataset is categorized into three distinct subsets: two private IC substrate datasets (Chip-side and Ball-side) and a patterned wafer dataset. To support high-efficiency training and pixel-wise localization, the images are provided as cropped samples in 256\u00d7256 and 200\u00d7200 sizes. Structurally, the data is organized into a pair_set for direct comparison tasks and a single_set for individual image analysis, covering both training and testing phases."

半导体异常检测（Semiconductor Anomaly Detection，简称Semi-AD）数据集是面向半导体制造领域基于对比的无监督异常检测（Anomaly Detection，简称AD）任务的专用基准数据集。该数据集针对复杂电路图案中罕见缺陷识别这一关键挑战，通过利用半导体器件固有的重复结构予以解决。数据集提供由参考器件与待测器件组成的精准对齐图像对，缺陷区域的真值标注（ground truth）由对齐图像对之间的逐像素差异计算得到。该数据集共分为三个独立子集：两个私有集成电路（Integrated Circuit, IC）衬底数据集（芯片侧与焊球侧），以及一个图案化晶圆数据集。为支撑高效训练与逐像素定位任务，数据集提供尺寸为256×256与200×200的裁剪图像样本。从数据组织形式来看，数据集分为用于直接对比任务的pair_set与用于单张图像分析的single_set，覆盖训练与测试全流程。