Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PUAD: Frustratingly Simple Method for Robust Anomaly Detection

Shota Sugawara, Ryuji Imamura

TL;DR

This work tackles robust anomaly detection in industrial imagery by distinguishing between picturable and unpicturable anomalies, arguing that reconstruction-based methods struggle with logical anomalies. It introduces PUAD, which uses EfficientAD for picturable anomalies and a fast feature-based Mahalanobis detector for unpicturable anomalies, with separate normalization before summation. The approach yields state-of-the-art AUROC on the MVTec LOCO AD dataset while incurring only a small latency increase, demonstrating the value of tailoring methods to the nature of the anomaly. The findings suggest a practical direction for real-time anomaly detection: deploy specialized detectors for different anomaly types to balance accuracy and speed in industrial applications.

Abstract

Developing an accurate and fast anomaly detection model is an important task in real-time computer vision applications. There has been much research to develop a single model that detects either structural or logical anomalies, which are inherently distinct. The majority of the existing approaches implicitly assume that the anomaly can be represented by identifying the anomalous location. However, we argue that logical anomalies, such as the wrong number of objects, can not be well-represented by the spatial feature maps and require an alternative approach. In addition, we focused on the possibility of detecting logical anomalies by using an out-of-distribution detection approach on the feature space, which aggregates the spatial information of the feature map. As a demonstration, we propose a method that incorporates a simple out-of-distribution detection method on the feature space against state-of-the-art reconstruction-based approaches. Despite the simplicity of our proposal, our method PUAD (Picturable and Unpicturable Anomaly Detection) achieves state-of-the-art performance on the MVTec LOCO AD dataset.

PUAD: Frustratingly Simple Method for Robust Anomaly Detection

TL;DR

This work tackles robust anomaly detection in industrial imagery by distinguishing between picturable and unpicturable anomalies, arguing that reconstruction-based methods struggle with logical anomalies. It introduces PUAD, which uses EfficientAD for picturable anomalies and a fast feature-based Mahalanobis detector for unpicturable anomalies, with separate normalization before summation. The approach yields state-of-the-art AUROC on the MVTec LOCO AD dataset while incurring only a small latency increase, demonstrating the value of tailoring methods to the nature of the anomaly. The findings suggest a practical direction for real-time anomaly detection: deploy specialized detectors for different anomaly types to balance accuracy and speed in industrial applications.

Abstract

Developing an accurate and fast anomaly detection model is an important task in real-time computer vision applications. There has been much research to develop a single model that detects either structural or logical anomalies, which are inherently distinct. The majority of the existing approaches implicitly assume that the anomaly can be represented by identifying the anomalous location. However, we argue that logical anomalies, such as the wrong number of objects, can not be well-represented by the spatial feature maps and require an alternative approach. In addition, we focused on the possibility of detecting logical anomalies by using an out-of-distribution detection approach on the feature space, which aggregates the spatial information of the feature map. As a demonstration, we propose a method that incorporates a simple out-of-distribution detection method on the feature space against state-of-the-art reconstruction-based approaches. Despite the simplicity of our proposal, our method PUAD (Picturable and Unpicturable Anomaly Detection) achieves state-of-the-art performance on the MVTec LOCO AD dataset.
Paper Structure (11 sections, 2 equations, 4 figures, 3 tables)

This paper contains 11 sections, 2 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Anomaly Detection Tasks
  4. Anomaly Detection Methods
  5. Proposed Method
  6. Picturable and Unpicturable Anomalies
  7. Picturable Anomaly Detection
  8. Unpicturable Anomaly Detection
  9. Anomaly Score Normalization
  10. Experiments
  11. Conclusion

Figures (4)

  • Figure 1: These images from MVTec LOCO Dataset are examples of picturable and unpicturable anomalies. The white areas in the right column represent the Ground Truth, indicating the locations where anomalies are present. In the image of unpicturable anomaly, the entire bag is marked as the ground truth, which means that the location of the anomaly is not clearly specified.
  • Figure 2: Normal images, anomalous images, ground truth maps, and anomaly maps of EfficientAD for picturable and unpicturable anomalies. In unpicturable anomaly cases, the ground truth identifies all possible areas as candidates for the anomaly since it is impossible to identify a specific point of anomaly on the anomaly map. However, EfficientAD has not been able to produce the expected detection results.
  • Figure 3: The scheme of the proposed method. We consider the score from EfficientAD as the picturable anomaly score. In addition, we consider the Mahalanobis distance calculated using the student's former half output or the teacher's output as the unpicturable anomaly score. The final anomaly score is obtained by normalizing and summing up each score.
  • Figure 4: (a) An original image in the screw bag category of the MVTec LOCO Dataset. (b) An autoencoder's output image in the screw bag category of the MVTec LOCO Dataset. This shows that most features of logical information have been lost.