2026-10-11T08:32:59Z https://www.scidb.cn/oai

10.57760/sciencedb.010x0 2026-09-07T09:46:20Z

2026-09-07 LGPSD-DET doi:10.57760/sciencedb.010x0 en LGPSD-DET has collected a large number of light guide plate surface defects, containing 17,784 high-quality images with annotations for seven typical industrial defects: scratch (SCR), white stain (WST), scar (SCA), dark spot (DSP), white shadow (WSH), black shadow (BSH), and missing (MIS). The dataset covers imaging scenarios under varied illumination conditions and imaging perspectives, which comprehensively and realistically reflects the diversity and complexity of light guide plate surface defects in real manufacturing environments. The dataset is split into training, validation, and testing subsets following a two-stage instance-level stratified splitting strategy with a 7:1:2 ratio, which prevents multi-view images of the same physical defect from being distributed across different subsets and eliminates data leakage. Original defect images of the three subsets are packaged in the `images.zip` compressed folder, and corresponding TXT annotation files are stored in the `labels.zip` compressed folder. Besides, a `classes.txt` file enumerates all defect categories in a fixed order, which provides a mapping reference between category indices in annotation files and defect names. This dataset also includes all custom code supporting the findings of this work, including annotation format conversion scripts (YOLO to COCO and YOLO to VOC) as well as training and inference code for benchmark models. All supporting code is packaged in the `LGPSD-DET-code.zip` compressed folder. LGPSD-DET; light guide plate; defect detection Hanzheng Bao Shuo Peng Junfeng Li EMBARGO https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank