StructuralAD

Structural Knowledge-Based Anomaly Detection to Inspect Ball-Based Lens Actuators

https://doi.org/10.1109/ACCESS.2025.3622686

IEEE Access 2025, 13, 184110-184121
ISSN:2169-3536
eISSN:2169-3536

Repo Structure

The repository is organized as follows:

• custom/: custom defined module
  • detector: proposal Structural AD model
• datasets/: dataset folder
  • BALL: Structural AD datasets; https://doi.org/10.21227/mysd-9b62
    • ball: default datasets
    • enhanced_ball: Augmented dataset
• results/: results storage folder
• anomaly_detection.ipynb: training and testing all models
• img_enhancing.ipynb: Create enhanced image datasets enhanced_ball for detector

BALL Dataset

Image of faulty modules target to ball missing inspection collected at the optical cross-section inspection site:

Dataset	Task	Size	Normal	Anomalies
BaLL	Anomaly Detection	2352 x 2944	10,183	1,755

Example

from custom.data import Ball, MVTec
from custom.models import Patchcore, EfficientAd, Detector
from custom.engine import Engine

options = {}
method = 'Detector' # EfficientAd Patchcore EfficientAd Dinomaly
if method != 'Patchcore':
    options['train_batch_size'] = 1

category = 'ball' # enhanced_ball
if 'ball' in category:
    options['category'] = category

datamodule = Ball(**options) if category == 'ball' or category == 'enhanced_ball' else MVTec(**options)
model = Patchcore() if method == 'Patchcore' else EfficientAd() if method == 'EfficientAd' else Detector

engine = None
if method == 'patchcore':
    engine = Engine()
elif method == 'Detector':
    engine = Engine(is_detector=True)
elif method == 'EfficientAd':
    engine = Engine(max_epochs=10)

engine.fit(datamodule=datamodule, model=model)

if 'ball' in category:
    datasets_name = 'Ball'
else:
    datasets_name = 'Mvtec'
    
predictions = engine.predict(
    datamodule=datamodule,
    model=model,
    ckpt_path=f'./results/{method}/{datasets_name}/{category}/lastest/weights/lightning/model.ckpt',
    return_predictions=True,
)

engine.report()

* To run the code, you need to have Python 3.10.

Citation

If you use our code, please cite the paper below:

@article{jeon2025structural,
  title={Structural Knowledge-based Anomaly Detection to inspect Ball-based Lens Actuators},
  author={Jeon, Janggun and Ahn, Junho and Kim, Namgi},
  journal={IEEE Access},
  volume={13},
  number={},
  pages={184110-184121},
  year={2025},
  publisher={IEEE}
}