filter-inspection-data
收藏资源简介:
该数据集为IR-CUT滤光片裁片的暗场成像图,用于光学滤光片外观质检。数据来自11个采集包,包含3728张PNG图像和1977个labelme JSON格式的多边形标注(label字段恒为defect,缺陷类型由文件夹名决定)。图像涵盖良品(1746张)和多种缺陷:划痕(1170张)、崩边(565张)、麻点(164张)、碎(64张)、不可擦脏(19张)。数据集提供两种切分方案:splitA为全混合8:1:1随机切分(训练2768张、验证434张、测试409张),splitB为跨批次留包切分(测试集为样本_38016_P08262014整包)。分组单位以(包,session)为准,确保同一次拍摄不跨split。此外,数据集包含inventory.csv全量清点文件和metrics_fixed.csv几何指标文件。已知数据质量问题包括:约3.3%良品标签存疑(漏标缺陷)、34张缺陷图无监督标注、崩/侧面崩/碎缺陷标注多边形几乎全在片体外、麻点标注口径不统一、标注精细度分两档、采集配置色调漂移、判定规格缺失等。该数据集适用于图像分类、图像分割、工业缺陷检测等任务。
This dataset consists of dark-field imaging images of IR-CUT filter slices, used for optical filter appearance quality inspection. The data comes from 11 acquisition packages, containing 3728 PNG images and 1977 polygon annotations in labelme JSON format (the label field is always defect, and the defect type is determined by the folder name). The images cover good products (1746 images) and various defects: scratches (1170), chipped edges (565), pitting (164), broken (64), and non-erasable dirt (19). The dataset provides two splitting schemes: splitA is a fully mixed 8:1:1 random split (2768 training, 434 validation, 409 test), and splitB is a cross-batch leave-package split (test set is the entire package of sample_38016_P08262014). Grouping is based on (package, session) to ensure that the same shoot is not split across splits. Additionally, the dataset includes an inventory.csv full inventory file and a metrics_fixed.csv geometric metrics file. Known data quality issues include: about 3.3% of good product labels are questionable (missing defect labels), 34 defect images have no supervised annotations, the annotation polygons for chipped/side chipped/broken defects are almost entirely outside the slice body, inconsistent annotation standards for pitting, two levels of annotation granularity, color shift in acquisition configuration, and missing judgment specifications. The dataset is suitable for tasks such as image classification, image segmentation, and industrial defect detection.
滤光片外观质检数据集
该数据集为IR-CUT 滤光片裁片的暗场成像图,来源于 11 个采集包(样本7_27.zip),并附带配套模型与开箱即用推理包。
数据规模
| 类别 | 张数 |
|---|---|
| 良品 | 1746 |
| 划痕(划) | 1170 |
| 崩边(崩 + 侧面崩) | 565 |
| 麻点(麻) | 164 |
| 碎 | 64 |
| 不可擦脏 | 19 |
| 合计 | 3728 |
标注文件为 1977 个 labelme JSON(多边形),其中 label 字段恒为 defect,缺陷类型由文件夹名决定。
文件结构
images/:3728 张 PNG + 1977 个 labelme JSON,保持原始 11 包目录结构splits/splitA/meta.csv:全混合 8:1:1 切分(train 2768 / val 434 / test 409)splits/splitB/meta.csv:跨批次留包切分(test = 样本_38016_P08262014 整包)splits/*/dropped.csv:被剔除的图与原因inventory.csv:3728 张全量清点(批次/机器/工单/标注粒度/顶点数等)metrics_fixed.csv:1948 张已标注图的几何指标(顶点密度、片体内比例等)
切分协议
- 分组单位为 (包, session),同一次拍摄的图绝不跨 split
- 两套切分均实测 0 组跨 split(共 288 组)
test为锁箱,训练与选型全程不加载
命名规律
样本_38016_P08262014 → 批次 38016 / 机器 P08 / 工单 262014。路径中缺失的信息一律留空,未做任何推断——5 个包无机器号,2 个连批次号也没有。
已知数据质量问题
- 良品标签约 3.3% 存疑(漏标的真缺陷),已用客观门槛核查
- 34 张缺陷图零监督(6 张缺 json + 28 张 shapes 为空),已在
dropped.csv标出 - 崩/侧面崩/碎 的标注多边形几乎整个画在片体外(片体内比例仅 1–3%),语义上成立但直接栅格化做分割 GT 会有问题
- 麻点标注口径不统一:多数包逐颗描(面积中位 0.15–0.64% 图幅),但
样本_38030_P15261973_2是画大圈圈住密集区(中位 9.42%),差 20–60 倍 - 标注精细度分两档:labelme 3.3.10 的包顶点中位 80(精描),5.11.4 的包顶点中位 12(粗框),
样本_38031两版本混装 - 采集配置色调漂移:B 通道逐日 15→8→5,冷/暖两种配置
- 判定规格缺失:多亮×多大的圆点算麻点、多长的划痕算不良,至今无书面判据
相关文档
docs/INDEX.md:项目总索引(四个仓库怎么选、三个核心结论、数据一览)docs/REPRODUCE.md:完整复现指南(环境、数据准备、逐个实验的命令与预期数字)docs/FILES.md:逐文件使用说明(41 个脚本、结果表字段、51 个训练产物目录)docs/RESULTS.md:全部实验结果(42 个模型完整榜单、分割 vs 分类同口径对决)docs/AUDIT.md:诚实性审计(四道审计的原理与结果)docs/CAVEATS.md:已知问题与边界(含评估口径错误)docs/REPOS.md:四个仓库逐一说明与下载方式
配套模型与推理包:https://huggingface.co/daipath/filter-inspection




