Self-Supervised Masked Convolutional Transformer Block for Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Anomaly detection is crucial in computer vision for tasks like fault detection, surveillance, and medical image analysis.
• Current methods often frame anomaly detection as one-class classification using normal data only.
• Reconstruction-based methods indicate abnormality by reconstruction error of masked normal inputs.

Purpose of the Study:

• To introduce a novel self-supervised masked convolutional transformer block (SSMCTB) for enhanced anomaly detection.
• To demonstrate the flexibility and broad applicability of the proposed block across various domains and neural architectures.
• To improve upon existing reconstruction-based anomaly detection techniques.

Main Methods:

• Developed a novel self-supervised masked convolutional transformer block (SSMCTB).
• Integrated a 3D masked convolutional layer, transformer for channel-wise attention, and a Huber loss objective.
• Extended previous SSPCAB with enhanced self-supervised learning capabilities.
• Applied SSMCTB to multiple state-of-the-art neural models for anomaly detection.

Main Results:

• The SSMCTB demonstrated considerable performance improvements on five diverse benchmarks: MVTec AD, BRATS, Avenue, ShanghaiTech, and Thermal Rare Event.
• Showcased applicability to anomaly detection in medical images and thermal videos, in addition to RGB and surveillance videos.
• Empirical results confirmed the generality and flexibility of the proposed block.

Conclusions:

• The novel SSMCTB offers a flexible and effective approach to anomaly detection in computer vision.
• The block's architecture and self-supervised objective lead to significant performance gains across various tasks and datasets.
• SSMCTB represents a flexible, core architectural component applicable to a wide range of neural networks for anomaly detection.