In Generalizable Few-Shot Anomaly Detection (GFSAD), a common model must be learned and shared across several categories, while simultaneously ensuring that the model is adaptable to new categories with a restricted number of normal images. While CNN-transformer architectures obtain high success in many vision tasks, the potential of CNN-transformer architectures in GFSAD is still to be discovered. In this paper, we introduce ADFormer, a dual CNN-transformer architecture that combines the strengths of CNNs and transformers, with the aim of learning discriminative features that have both local and global receptive fields. We also incorporate a self-supervised bipartite matching approach in ADFormer that reconstructs query images from support images, followed by detecting anomalies based on the high loss in reconstruction. Additionally, we present a consistency-enhanced loss to enhance the spatial and semantic consistency of features, thereby reducing the dependence on a large AD dataset for training. Experimental results show that ADFormer with consistency-enhanced loss significantly improves GFSAD performance. Compared to other anomaly detection methods, ADFormer outperforms considerably on the MVTec AD, MPDD, and VisA datasets.
We follow RegAD for the dataset preparation.
Download the support dataset for few-shot anomaly detection on Google Drive or Baidu Disk (i9rx) and unzip the dataset. For those who have problem downloading the support set, please optional download categories of capsule and grid on Baidu Disk (pll9) and Baidu Disk (ns0n).
python train.pypython test.py- Thanks to MaskFormer and RegAD for their wonderful work and code!
@article{zhu2024adformer,
title={ADFormer: Generalizable Few-Shot Anomaly Detection with Dual CNN-Transformer Architecture},
author={Zhu, Bingke and Gu, Zhaopeng and Zhu, Guibo and Chen, Yingying and Tang, Ming and Wang, Jinqiao},
journal={IEEE Transactions on Instrumentation and Measurement},
year={2024},
publisher={IEEE}
}