Target Before Shooting: Accurate Anomaly Detection and Localization Under One Millisecond via Cascade Patch Retrieval

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Anomaly Detection (AD) is crucial for identifying defects and unusual patterns.
• Existing patch-matching methods for AD face challenges in efficiency and accuracy.
• Re-examining the matching process in AD is key to developing improved frameworks.

Purpose of the Study:

• To propose a novel Anomaly Detection (AD) framework named Cascade Patch Retrieval (CPR).
• To achieve both state-of-the-art accuracy and high running speed in anomaly detection.
• To provide an efficient and effective solution for identifying anomalies in images.

Main Methods:

• A cascade patch retrieval procedure is employed for anomaly detection.
• Top-K similar training images are selected using robust histogram matching.
• Nearest neighbors for test image patches are retrieved using a trained local metric and non-background probability.

Main Results:

• The proposed Cascade Patch Retrieval (CPR) method consistently outperforms existing state-of-the-art (SOTA) methods on MVTec AD, BTAD, and MVTec-3D AD datasets.
• CPR achieves record-breaking accuracy in anomaly detection.
• CPR demonstrates exceptional efficiency, running at 113 FPS with a simplified version processing images in under 1 ms.

Conclusions:

• Cascade Patch Retrieval (CPR) offers a significant advancement in anomaly detection.
• The framework provides a superior balance of accuracy and speed compared to current methods.
• CPR is a highly efficient and effective solution for real-world anomaly detection applications.