Masked Attention ConvNeXt Unet with Multi-Synthesis Dynamic Weighting for Anomaly Detection and Localization


SHIH CHIH LIN (National Tsing Hua University),* Ho Weng Lee (National Tsing Hua University), Yu-Shuan Hsieh (National Tsing Hua University), Cheng Yu Ho (National Tsing Hua University), Shang-Hong Lai (National Tsing Hua University)
The 34th British Machine Vision Conference

Abstract

In recent years, self-supervised models like Cutpaste, Mask, NSA, and Perlin have gained popularity for anomaly detection. These models generate synthetic data by employing various data augmentation strategies, demonstrating their potential for improving anomaly detection through learned representations. In this study, we introduce an algorithm called Multi-Synthesis Dynamic Weighting (MSdW) to leverage the advantages of diverse synthetic data. MSdW enables the model to learn various abnormal conditions during training, thereby enhancing accuracy. Our model architecture consists of reconstructive and discriminative subnetworks, both utilizing the UNet architecture. The encoders in both subnetworks employ modern ConvNets, specifically ConvNeXtV2, for proficient feature extraction. Additionally, we propose an attention mechanism known as Self-Supervised Predictive Convolutional Block with Multi-Attentions (SSPCBMA), which is seamlessly integrated into the reconstructive subnetwork to enhance feature extraction capabilities. We evaluate our proposed model on multiple datasets designed for anomaly detection and segmentation tasks, including MVTecAD, BTAD, and KSDD2. These datasets serve various purposes, and our model outperforms the state-of-the-art methods, particularly in terms of Pixel AP and PRO indices.

Video



Citation

@inproceedings{LIN_2023_BMVC,
author    = {SHIH CHIH LIN and Ho Weng Lee and Yu-Shuan Hsieh and Cheng Yu Ho and Shang-Hong Lai},
title     = {Masked Attention ConvNeXt Unet with Multi-Synthesis Dynamic Weighting for Anomaly Detection and Localization},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year      = {2023},
url       = {https://papers.bmvc2023.org/0911.pdf}
}


Copyright © 2023 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