multi-class anomaly detection is more efficient and less resource-consuming in industrial anomalydetection scenes that involve multiple categories or exhibit large intra-class diversity. However, most industrial imag...
详细信息
multi-class anomaly detection is more efficient and less resource-consuming in industrial anomalydetection scenes that involve multiple categories or exhibit large intra-class diversity. However, most industrial image anomalydetection methods are developed for one-classanomalydetection, which typically suffer significant performance drops in multi-class scenarios. Research specifically targeting multi-class anomaly detection remains relatively limited. In this work, we propose a powerful unified normalizing flow for multi-class anomaly detection, which we call UniFlow. A multi-cognitive visual adapter (Mona) is employed in our method as the feature adaptation layer to adapt image features for both the multi-class anomaly detection task and the normalizing flow model, facilitating the learning of general knowledge of normal images across multiple categories. We adopt multi-cognitive convolutional networks with high capacity to construct the coupling layers within the normalizing flow model for more effective multi-class distribution modeling. In addition, we employ a multi-scale feature fusion module to aggregate features from various levels, thereby obtaining fused features with enhanced expressive capabilities. UniFlow achieves a class-average image-level AUROC of 99.1% and a class-average pixel-level AUROC of 98.0% on MVTec AD, outperforming the SOTA multi-class anomaly detection methods. Extensive experiments on three benchmark datasets, MVTec AD, VisA, and BTAD, demonstrate the efficacy and superiority of our unified normalizing flow in multi-class anomaly detection.
Existing unsupervised multi-class anomaly detection algorithms usually train unified reconstruction networks to capture the distribution of all classes simultaneously. However, under such a challenging setting, popula...
详细信息
Existing unsupervised multi-class anomaly detection algorithms usually train unified reconstruction networks to capture the distribution of all classes simultaneously. However, under such a challenging setting, popular reconstruction networks need to be elaborately designed to avoid the "identical shortcut''. In addition, the distribution of each category is different, which may mean different requests for expression ability. To solve these problems, built on the intuitive "classification-then-detection"idea, we utilize clustering algorithm to expose the category information hidden in the pre-trained deep features, then propose a simple and application- friendly approach for multi-class anomaly detection. The proposed approach consists of Category Anchor Construction (CAC), Category Information Mining (CIM) and Local Feature Routing (LFR). Firstly, CAC is proposed to extract the corresponding pre-trained features from a small subset of training images to construct category anchors, preserving the valuable category information provided by the training set. Then, CIM is introduced to mine category information embedded in pre-trained features by category anchors voting and acquires the category labels. Finally, to achieve multi-class anomaly detection, we propose LFR, splitting multi-class distribution into multiple single-class distributions according to category labels so that separate single-classanomalydetection heads can be trained to express them. In spite of simplicity, the proposed method outperforms state-of-the-art algorithms in terms of accuracy and stability on the widely used MVTec-AD, VisA, MVTec-LOCO, MPDD and BTAD datasets.
This work studies a challenging and practical issue known as multi-class unsupervised anomalydetection (MUAD). This problem requires only normal images for training while simultaneously testing both normal and anomal...
详细信息
This work studies a challenging and practical issue known as multi-class unsupervised anomalydetection (MUAD). This problem requires only normal images for training while simultaneously testing both normal and anomaly images across multiple classes. Existing reconstruction-based methods typically adopt pyramidal networks as encoders and decoders to obtain multi-resolution features, often involving complex sub-modules with extensive handcraft engineering. In contrast, a plain Vision Transformer (ViT) showcasing amore straightforward architecture has proven effective in multiple domains, including detection and segmentation tasks. It is simpler, more effective, and elegant. Following this spirit, we explore the use of only plain ViT features for MUAD. We first abstract a Meta-AD concept by synthesizing current reconstruction-based methods. Subsequently, we instantiate a novel ViT-based ViTAD structure, designed incrementally from both global and local perspectives. This model provide a strong baseline to facilitate future research. Additionally, this paper uncovers several intriguing findings for further investigation. Finally, we comprehensively and fairly benchmark various approaches using seven metrics and their average. Utilizing a basic training regimen with only an MSE loss, ViTAD achieves state-of-the-art results and efficiency on MVTec AD, VisA, and Uni-Medical datasets. E.g., achieving 85.4 mAD that surpasses UniAD by +3.0 for the MVTec AD dataset, and it requires only 1.1 hand 2.3G GPU memory to complete model training on a single V100 that can serve as a strong baseline to facilitate the development of future research. Full code is available at https: //***/projects/ViTAD/.
The ability to recognize and model human Activities of Daily Living (ADL) and to detect possible deviations from regular patterns, or anomalies, constitutes an enabling technology for developing effective Socially Ass...
详细信息
ISBN:
(纸本)9783030052041;9783030052034
The ability to recognize and model human Activities of Daily Living (ADL) and to detect possible deviations from regular patterns, or anomalies, constitutes an enabling technology for developing effective Socially Assistive Robots. Traditional approaches aim at recognizing an anomaly behavior by means of machine-learning techniques trained on anomalies' dataset, like subject's falls. The main problem with these approaches lies in the difficulty to generate these dataset. In this work, we present a two-step framework implementing a new strategy for the detection of ADL anomalies. Indeed, rather than detecting anomaly behaviors, we aim at identifying those that are divergent from normal ones. This is achieved by a first step, where a deep learning technique determine the most probable ADL class related to the action performed by the subject. In a second step, a Gaussian Mixture Model is used to compute the likelihood that the action is normal or not, within that class. We performed an experimental validation of the proposed framework on a public dataset. Results are very close to the best traditional approaches, while at the same time offering the significant advantage that it is much easier to create dataset of normal ADL.
暂无评论