Annotated dataset for surface defect detection on ceramic substrates in manufacturing.

Ceramic substrates are critical foundational materials in semiconductor packaging and electronic device manufacturing. Surface defects on these substrates-such as adhesion, scar, spot, uneven, abnormal thickness, and edge defect-directly impact device performance and reliability. However, existing datasets commonly suffer from issues such as limited defect types, low image resolution, lack of realistic manufacturing process backgrounds, and incomplete annotations. To our knowledge, no publicly available dataset currently covers all six ceramic-substrate defect categories included in this study. This gap limits the development and generalizability of deep learning-based automated defect detection models. To fill this gap, this study constructs and releases a high-resolution, multi-defect-type dedicated dataset for ceramic substrate defect detection. The dataset comprises 2,892 high-resolution annotated images of ceramic substrates, encompassing six typical defect types. The image pixel range covers from 1024×1024 to 4096×4000 pixels, fully preserving the overall layout and subtle defect characteristics. The dataset contains a total of 7,689 defect instances, all visible defects are annotated with precise bounding boxes, and the annotation files follow the COCO JSON format. Conversion versions in PASCAL VOC and YOLO formats are also provided to ensure compatibility with mainstream object detection frameworks (e.g., YOLOv8, YOLOv11, Faster R-CNN, etc.). This dataset is designed to provide researchers in the fields of computer vision, Automated Optical Inspection (AOI), and semiconductor quality control with standardized, highly reusable benchmark resources, driving the development of ceramic substrate defect detection algorithms toward high precision and high robustness, thereby offering data support for the intelligent upgrade of the semiconductor manufacturing industry.