Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
110 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
44 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Patch SVDD: Patch-level SVDD for Anomaly Detection and Segmentation (2006.16067v2)

Published 29 Jun 2020 in cs.CV

Abstract: In this paper, we address the problem of image anomaly detection and segmentation. Anomaly detection involves making a binary decision as to whether an input image contains an anomaly, and anomaly segmentation aims to locate the anomaly on the pixel level. Support vector data description (SVDD) is a long-standing algorithm used for an anomaly detection, and we extend its deep learning variant to the patch-based method using self-supervised learning. This extension enables anomaly segmentation and improves detection performance. As a result, anomaly detection and segmentation performances measured in AUROC on MVTec AD dataset increased by 9.8% and 7.0%, respectively, compared to the previous state-of-the-art methods. Our results indicate the efficacy of the proposed method and its potential for industrial application. Detailed analysis of the proposed method offers insights regarding its behavior, and the code is available online.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (2)
  1. Jihun Yi (11 papers)
  2. Sungroh Yoon (163 papers)
Citations (353)

Summary

  • The paper introduces Patch SVDD, a novel patch-level method that integrates self-supervised learning with SVDD to localize anomalies.
  • The authors report performance gains of up to 9.8% in detection and 7.0% in segmentation on the MVTec AD dataset.
  • The approach offers scalable, precise anomaly analysis, setting a new benchmark for industrial quality control and real-time monitoring.

Patch SVDD: Patch-Level Support Vector Data Description for Anomaly Detection and Segmentation

The paper "Patch SVDD: Patch-level SVDD for Anomaly Detection and Segmentation" by Jihun Yi and Sungroh Yoon proposes a novel approach—Patch SVDD—aimed at improving the efficacy of anomaly detection and segmentation in images. The approach extends the deep learning variant of Support Vector Data Description (SVDD) to a patch-based method using self-supervised learning. This development shows promise in addressing challenges associated with industrial applications where detecting and localizing anomalies is critical.

Methodological Advances

Traditional anomaly detection focuses on a binary classification problem—a task ideally suited to one-class classification algorithms like One-class Support Vector Machine (OC-SVM) and SVDD. Ruff et al.'s Deep SVDD approach, which leverages deep neural networks, eliminated the need for pre-selecting a kernel function by training neural networks to extract data-dependent representations. Deep SVDD, however, is limited to global image analysis.

Patch SVDD differentiates itself by enabling localized inspection, thus addressing intra-class variation challenges by analyzing smaller image sections or patches. By utilizing self-supervised learning alongside a hierarchical encoding structure, the method permits multi-scale inspection, paving the way for detailed anomaly localization at the pixel level. The integration of self-supervised tasks, as articulated by Doersch et al., enhances the ability of neural networks to form meaningful representations without reliance on labeled data.

Empirical Evaluation

On the MVTec AD dataset, which contains 15 classes of industrial images, Patch SVDD demonstrates notable improvements over previous approaches. Anomaly detection and segmentation performances rose by 9.8% and 7.0%, respectively, measured by area under the receiver operating characteristic curve (AUROC). The various experimental configurations highlight the anomalies' multi-scale nature, with anomaly maps showcasing precise localization of defects. Importantly, self-supervised tasks contribute significantly to addressing the high intra-class variation, especially in object categories.

Insights into Random Encoders

Intriguingly, the paper explores the effectiveness of random encoders and nearest neighbor algorithms using raw patches, revealing their potential in certain classes where the network architecture alone drives separability. This insight suggests underlying simplicity in the data structure that could be leveraged in other unsupervised contexts.

Implications and Future Directions

Patch SVDD's strong performance heralds advancements in industrial quality control, predictive maintenance, and similar domains where anomalies must be detected and localized with high precision. Its computational efficiency and scalability are clear assets in practical settings, especially when combined with self-supervised learning techniques.

Future work could explore more complex self-supervised tasks or alternative architecture designs to further optimize performance. Additionally, an investigation into the interpretability of anomaly maps and their integration into real-time anomaly monitoring systems would be valuable, enhancing the method's applicability in active industrial environments.

Overall, Patch SVDD presents a robust, empirically validated methodology that enhances the granularity and accuracy of anomaly detection and segmentation, establishing a new benchmark for similar applications in computer vision and beyond.