Img2PatchSeqAD: Industrial Image Anomaly Detection Based on Image Patch Sequence

Abstract
In the domain of industrial Visual Anomaly Detection(VAD), methods based on image reconstruction are the most popular and successful approaches. However, current image reconstruction methods rely on global image information, which proves to be both blind and inefficient for anomaly detection tasks. Our approach tackles these limitations by taking advantage of neighboring image patches to assess the presence of anomalies in the current image and then selectively reconstructing those patches. In this paper, we introduce a novel architecture for image anomaly detection, named Img2PatchSeqAD. Specifically, we employ a row-wise scanning method to construct sequences of image patches and design a network framework based on an image patch sequence encoder-decoder structure. Additionally, we utilize the KAN model and ELA attention mechanism to develop methods for image patch vectorization and establish an image reconstruction pipeline. Experimental results on the MVTec-AD and VisA datasets demonstrate the effectiveness of our approach, achieving localization and detection scores of 81.3 (AUROC) and 91.9 (AP) on the multi-class MVTec-AD dataset.
Description

CCS Concepts: Computing methodologies->Image segmentation; Scene anomaly detection

        
@inproceedings{
10.2312:pg.20241324
, booktitle = {
Pacific Graphics Conference Papers and Posters
}, editor = {
Chen, Renjie
and
Ritschel, Tobias
and
Whiting, Emily
}, title = {{
Img2PatchSeqAD: Industrial Image Anomaly Detection Based on Image Patch Sequence
}}, author = {
Liu, Yang
and
Ji, Ya Tu
and
Wang, L.
and
Dai, L. J.
and
Yao, Miao Miao
and
Li, Xiao Mei
and
Xue, Xiang
and
Xu, H. T.
and
Ren, Qing Dao Er Ji
and
Shi, Bao
and
Wu, N. E.
and
Lu, M.
and
Xu, Xuan Xuan
and
Guo, H. X.
}, year = {
2024
}, publisher = {
The Eurographics Association
}, ISBN = {
978-3-03868-250-9
}, DOI = {
10.2312/pg.20241324
} }
Citation