Anomaly Detection Based on Semi-Formula Driven Pre-training Dataset to Represent Subtle Difference and Anomaly Score


Hiroki Kobayashi (Chukyo University), Naoki Murakami (Chukyo University), Naoto Hiramatsu (Chukyo University), Takahiro Suzuki (Chukyo University), Manabu Hashimoto (Chukyo University)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

The goal of a surface anomaly detection task is to classify an inspection image and pixel as normal/anomaly with high precision. A typical conventional method called PaDiM pre-trains convolutional neural networks with the ImageNet dataset for 1000-class classification and detects the anomaly images from deviance on the feature space. However, since a single class in ImageNet has a wide range of meanings, it is difficult to represent subtle difference between normal and anomaly images as different features. Moreover, PaDiM assumes that two images with similar anomaly scores have features with similar values. However, the feature space is made to classify the ImageNet class, it is not designed to assign two images with similar anomaly scores to relatively similar features. Therefore, we propose an anomaly detection method based on pre-training using novely semi-formula driven image dataset to represent “subtle difference” between two images as different features and two images with similar “anomaly score” as similar features. An image dataset for pre-training is generated by adding pseudo-defects with random Gaussian Mixture Model (GMM) parameters to an existing image dataset. GMM parameters have a different value for each parameter, but the appearances of the generated images have only subtle difference. Next, the regression network is pre-trained to estimate GMM parameters that represent the anomaly score of generated anomaly images. In the experiments with MVTecAD, the proposed method achieved high precision anomaly detection for categories where ImageNet performed poorly.

Citation

@inproceedings{Kobayashi_2024_BMVC,
author    = {Hiroki Kobayashi and Naoki Murakami and Naoto Hiramatsu and Takahiro Suzuki and Manabu Hashimoto},
title     = {Anomaly Detection Based on Semi-Formula Driven Pre-training Dataset to Represent Subtle Difference and Anomaly Score},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0833.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection