@article{ART003329475},
author={Ki-Ju Kim and Gye-Young Kim},
title={Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection},
journal={Journal of The Korea Society of Computer and Information},
issn={1598-849X},
year={2026},
volume={31},
number={4},
pages={97-106}
TY - JOUR
AU - Ki-Ju Kim
AU - Gye-Young Kim
TI - Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection
JO - Journal of The Korea Society of Computer and Information
PY - 2026
VL - 31
IS - 4
PB - The Korean Society Of Computer And Information
SP - 97
EP - 106
SN - 1598-849X
AB - Defective products can cause substantial economic losses and even threaten human safety; therefore, early defect detection prior to shipment is crucial. In conventional memory-bank-based approaches that model normal data, however, memory consumption increases sharply with the number of images and their resolution when all normal samples are used for modeling. In real industrial environments, large-scale and high-resolution images are often required for stable inspection, which limits practical deployment. To address this issue, we propose an online cluster-based memory bank. The proposed method processes normal images sequentially and models them in a memory-efficient manner by clustering their features online. Experiments on the Real-IAD dataset show that our method achieves an image-level AUROC of 90.6% while reducing memory usage by more than 90%. Furthermore, enabling high-resolution processing with 512×512 inputs improves the image-level AUROC by about 1.4 percentage points and increases the PRO metric by about 5 percentage points compared to the 224×224 baseline.
KW - Computer vision;Image processing;Industrial anomaly detection;Online clustering;Memory bank
DO -
UR -
ER -
Ki-Ju Kim and Gye-Young Kim. (2026). Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection. Journal of The Korea Society of Computer and Information, 31(4), 97-106.
Ki-Ju Kim and Gye-Young Kim. 2026, "Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection", Journal of The Korea Society of Computer and Information, vol.31, no.4 pp.97-106.
Ki-Ju Kim, Gye-Young Kim "Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection" Journal of The Korea Society of Computer and Information 31.4 pp.97-106 (2026) : 97.
Ki-Ju Kim, Gye-Young Kim. Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection. 2026; 31(4), 97-106.
Ki-Ju Kim and Gye-Young Kim. "Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection" Journal of The Korea Society of Computer and Information 31, no.4 (2026) : 97-106.
Ki-Ju Kim; Gye-Young Kim. Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection. Journal of The Korea Society of Computer and Information, 31(4), 97-106.
Ki-Ju Kim; Gye-Young Kim. Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection. Journal of The Korea Society of Computer and Information. 2026; 31(4) 97-106.
Ki-Ju Kim, Gye-Young Kim. Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection. 2026; 31(4), 97-106.
Ki-Ju Kim and Gye-Young Kim. "Memory-efficient Image Patch Distribution Modeling for Industrial Anomaly Detection" Journal of The Korea Society of Computer and Information 31, no.4 (2026) : 97-106.
