
摘要
在大规模工业制造中,能够识别缺陷部件是一项至关重要的任务。本文所解决的一个关键挑战是“冷启动问题”:仅使用正常(无缺陷)样本图像来训练模型。尽管针对每一类缺陷均可设计手工解决方案,但我们的目标是构建能够自动在多种不同任务上均表现优异的系统。目前表现最佳的方法是将ImageNet模型的特征嵌入与异常检测模型相结合。本文在此基础上进一步拓展,提出PatchCore,该方法利用一个最具代表性的正常样本块(patch)特征记忆库。PatchCore在保持优异推理速度的同时,实现了检测与定位任务的最先进性能。在具有挑战性且广泛使用的MVTec AD基准测试中,PatchCore的图像级异常检测AUROC得分最高达到99.6%,相比次优竞争对手的错误率降低了一半以上。此外,我们在两个额外的数据集上也取得了具有竞争力的结果,并在少样本场景下同样表现出色。\freefootnote{$^*$ 本工作为在亚马逊AWS实习期间完成。} 代码地址:github.com/amazon-research/patchcore-inspection。
代码仓库
captainfffsama/pathcore
pytorch
GitHub 中提及
OpenAOI/anodet
pytorch
GitHub 中提及
JoegameZhou/PatchCore
mindspore
GitHub 中提及
amazon-science/patchcore-inspection
pytorch
GitHub 中提及
yangyucheng000/patchcore
mindspore
any-tech/PatchCore-ex
pytorch
GitHub 中提及
Ultranity/Anomaly.Paddle
paddle
GitHub 中提及
totoroKalic/patchcore-mindspore
mindspore
GitHub 中提及
tiskw/patchcore-ad
pytorch
GitHub 中提及
taikiinoue45/PatchCore
pytorch
rvorias/ind_knn_ad
pytorch
openvinotoolkit/anomalib
pytorch
hcw-00/PatchCore_anomaly_detection
pytorch
GitHub 中提及
tbcvContributor/DeepHawkeye
pytorch
GitHub 中提及
amazon-research/patchcore-inspection
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| anomaly-classification-on-goodsad | PatchCore-100% | AUPR: 86.1 AUROC: 85.5 |
| anomaly-classification-on-goodsad | PatchCore-1% | AUPR: 83.3 AUROC: 81.4 |
| anomaly-detection-on-aebad-s | PatchCore | Detection AUROC: 71.0 Segmentation AUPRO: 87.8 |
| anomaly-detection-on-aebad-v | PatchCore | Detection AUROC: 70.7 |
| anomaly-detection-on-mpdd | PatchCore | Detection AUROC: 82.12 Segmentation AUROC: 95.66 |
| anomaly-detection-on-mvtec-ad | PatchCore Large | Detection AUROC: 99.6 FPS: 5.88 Segmentation AUPRO: 93.5 Segmentation AUROC: 98.2 |
| anomaly-detection-on-mvtec-ad | PatchCore | Detection AUROC: 99.2 Segmentation AUROC: 98.4 |
| anomaly-detection-on-mvtec-ad | PatchCore(16shot) | Detection AUROC: 95.4 |
| anomaly-detection-on-mvtec-loco-ad | PatchCore | Avg. Detection AUROC: 80.3 Detection AUROC (only logical): 75.8 Segmentation AU-sPRO (until FPR 5%): 39.7 |
| anomaly-detection-on-mvtec-loco-ad | PatchCore Ensemble | Avg. Detection AUROC: 79.4 Detection AUROC (only logical): 71.0 Detection AUROC (only structural): 87.7 Segmentation AU-sPRO (until FPR 5%): 36.5 |