Introducing Fast Fourier Convolutions into Anomaly Detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Anomaly detection is challenging due to anomalies appearing at test time.
• Reconstruction-based methods using Convolutional Neural Networks (CNNs) have limitations in capturing global information and precise localization due to local receptive fields.

Purpose of the Study:

• To propose FFC-AD, a novel reconstruction framework for anomaly detection.
• To enhance discrimination and localization capabilities by incorporating global information early in the detection process.
• To mitigate overgeneralization in anomaly detection models.

Main Methods:

• Utilizing Fourier Feature Convolutions (FFCs) within a reconstruction framework to capture long-range dependencies.
• Introducing Hidden Space Anomaly Simulation (HSAS) as a latent-space regularization technique.
• Evaluating the proposed FFC-AD framework on established anomaly detection datasets like MVTec AD and VisA.

Main Results:

• FFC-AD demonstrates significant improvements in anomaly detection accuracy.
• The method achieves superior performance in anomaly segmentation compared to existing state-of-the-art approaches.
• FFCs effectively capture global information, enhancing feature discrimination.

Conclusions:

• FFC-AD offers a robust solution for anomaly detection by integrating global context and mitigating overgeneralization.
• The proposed HSAS strategy effectively regularizes the latent space, improving model generalization.
• FFC-AD represents a significant advancement in reconstruction-based anomaly detection, particularly for complex visual inspection tasks.