Kernekoncepter
LTAD combines anomaly detection by reconstruction and semantic anomaly detection to detect defects across multiple and long-tailed image classes, without relying on dataset class names. It learns pseudo-class names and uses a VAE-based data augmentation to address the long-tailed distribution of real-world applications.
Resumé
The content discusses the problem of long-tailed anomaly detection (AD), where different image classes have vastly different sample sizes in real-world applications. It introduces several long-tailed AD datasets and performance metrics, and proposes a novel method called LTAD to address this challenge.
LTAD combines two approaches:
- AD by reconstruction: A transformer-based reconstruction module projects image patches onto the manifold of normal images, and the reconstruction error is used as an anomaly score.
- Semantic AD: A binary classifier in the semantic space of a pretrained foundation model (ALIGN) is used to detect anomalies, leveraging learned pseudo-class names to make the classifier class-sensitive.
To address the long-tailed distribution, LTAD has a two-phase training process:
- Phase 1 learns the pseudo-class names and a VAE-based data augmentation module to synthesize features for minority classes.
- Phase 2 then trains the reconstruction and classification modules using a mix of real and synthetic data.
Extensive experiments show that LTAD outperforms state-of-the-art methods on various long-tailed AD datasets and configurations. The ablation studies confirm the efficacy of LTAD's components, including the semantic AD module and the data augmentation strategy.
Statistik
"Anomaly detection aims to identify defective images and localize their defects (if any)."
"Various methods have shown that this problem can be solved with high accuracy; e.g., [3, 20, 31, 36, 41, 44, 69, 77, 79, 84] have success rates >95% for anomaly detection and localization on the MVTec dataset [5]."
"However, as illustrated in Fig. 1, these methods require a different model per image category, which compromises scalability to many classes."
Citater
"Anomaly detection (AD) aims to identify defective images and localize their defects (if any)."
"Ideally, AD models should be able to detect defects over many image classes; without relying on hard-coded class names that can be uninformative or inconsistent across datasets; learn without anomaly supervision; and be robust to the long-tailed distributions of real-world applications."