Computer Vision
双层DINOv3异常分数融合
Dual-Layer DINOv3 Anomaly Score FusionA method for unsupervised industrial anomaly detection that constructs separate memory banks from different transformer blocks of the DINOv3 model to retrieve distinct normal references, then combines their normalized anomaly maps with fixed weights to localize defects. / 一种用于无监督工业异常检测的方法，该方法从DINOv3模型的不同Transformer块构建独立的记忆库以检索不同的正常参考，然后使用固定权重组合其归一化的异常图以定位缺陷。
- Knowledge type
- 工艺技术
- First proposed or discovered
- 2026
- Proposer
- Bin Liang, Wei Liang, Haihua Liang, Qingfeng Jiang, Tao Zhang
- Creator
- 农场主
- Last updated
- 2026-11-06 23:06:11
Sub-knowledge points
Unsupervised anomaly detectionPixel-level defect localizationMemory bank constructionAnomaly map fusionDINOv2 comparisonPaDiM baseline
Higher-order knowledge points
Anomaly Score FusionDINOv3Unsupervised LearningIndustrial Anomaly Detection
Main parameters and performance indicators
DINOv3: Backbone model for feature extractionFourth-to-last transformer block: First feature layer for memory bankLast transformer block: Second feature layer for memory bankFixed equal weights: Combination weights for anomaly mapsMVTec AD: Benchmark dataset for evaluationVisA: Benchmark dataset for evaluationMPDD: Supplementary benchmark dataset for evaluation
Open problems
Image-wise normalization can weaken image-level detectionSeparate retrieval increases inference time
Related literature
ranked by community starsUnsupervised Industrial Anomaly Detection and Localization via Anomaly Score Fusion of Dual-Layer DINOv3 FeaturesOpen source
Related products and links
No item recorded yet
Creation and modification trail
1 recordsRecord title: Dual-Layer DINOv3 Anomaly Score Fusion.
Node update feedback
review required