[2025]
Few-shot anomaly detection aims to identify and locate image anomalies using limited normal samples, enabling rapid deployment in industrial and safety applications. However, recent multimodal pre-trained methods often rely on statistical comparisons between query and reference features, which may lead to overlapping anomaly scores due to neglecting normal region interference and local contextual correlations. To address this, we propose the Feature Discrepancy guided Score Learning (FDSL) method. FDSL explicitly models deviations from normal patterns via a Feature Discrepancy (FD) module that separates patch features into normal and deviation components. An Anomaly Score Learning (ASL) module is then introduced to capture feature correlations and integrate neighboring patch information for more discriminative scores. Additionally, a Local-enhanced Feature Extraction (LFE) mechanism improves detection of small-area anomalies, while an image feature index memory bank reduces redundancy to optimize memory usage. Experimental results on MVTec and VisA datasets show that FDSL achieves state-of-the-art performance in 1-shot, 2-shot, and 4-shot segmentation tasks and is competitive in classification, demonstrating its effectiveness and adaptability in few-sample scenarios.
Overview of FDSL. The framework consists of three core modules: the Feature Discrepancy (FD) module, the Anomaly Score Learning (ASL) module, and the Local-enhanced Feature Extraction (LFE) mechanism. During training, the FD module creates differential feature maps from intermediate features, which the ASL module refines into anomaly score maps using ground truth for optimization. During testing, the LFE mechanism improves detection of small area anomalies, and the outputs from both FD and ASL are merged to form the final anomaly score map.
MVTec: click here
VisA: click here
Taking the MVTec dataset as an example, generate the meta.json file.
Run the mvtec.py file provided in the data folder to generate the corresponding meta.json file.
cd data
python mvtec.pyStructure of MVTec Folder:
mvtec/
│
├── meta.json
│
├── bottle/
│ ├── ground_truth/
│ │ ├── broken_large/
│ │ │ └── 000_mask.png
| | | └── ...
│ │ └── ...
│ └── test/
│ ├── broken_large/
│ │ └── 000.png
| | └── ...
│ └── ...
│
└── ...
Apply the same approach to the VisA dataset.
Quick train and test
run trainAndTest.sh , if winmode=nan means don't use LFE, winmode=mul means use LFE
The parameters' meaning:
model_name: your model name for train/test results
device: cpu, cuda number
useNAM: use "Neighbor Aggregation Module" ?
winmode: option: nan, mul, onlywin
win-size: default: 13
win-stride: default: 2
win-weight: default: 0.5
seed: random seed
kshots: k-shots: 1, 2, 4
bash trainAndTest.sh {model_name} {device} {useNAM} {winmode} {win-size} {win-stride} {win-weight} {seed} {kshots}for example:
FDSL with NAM and LFE
bash trainAndTest.sh fdsl 0 1 mul 13 2 0.5 14 1FDSL with NAM and without LFE
bash trainAndTest.sh fdsl 0 1 nan 13 2 0.5 14 1FDSL without NAM and without LFE:
bash trainAndTest.sh fdsl 0 0 nan 13 2 0.5 14 1bash test_few_shot.sh {model_name} {device} {useNAM} {winmode} {win-size} {win-stride} {win-weight} {seed} {kshots}We thank for the code repository: open_clip, VAND, AnomalyCLIP, and WinCLIP.