Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection (2111.09099v6)

Abstract: Anomaly detection is commonly pursued as a one-class classification problem, where models can only learn from normal training samples, while being evaluated on both normal and abnormal test samples. Among the successful approaches for anomaly detection, a distinguished category of methods relies on predicting masked information (e.g. patches, future frames, etc.) and leveraging the reconstruction error with respect to the masked information as an abnormality score. Different from related methods, we propose to integrate the reconstruction-based functionality into a novel self-supervised predictive architectural building block. The proposed self-supervised block is generic and can easily be incorporated into various state-of-the-art anomaly detection methods. Our block starts with a convolutional layer with dilated filters, where the center area of the receptive field is masked. The resulting activation maps are passed through a channel attention module. Our block is equipped with a loss that minimizes the reconstruction error with respect to the masked area in the receptive field. We demonstrate the generality of our block by integrating it into several state-of-the-art frameworks for anomaly detection on image and video, providing empirical evidence that shows considerable performance improvements on MVTec AD, Avenue, and ShanghaiTech. We release our code as open source at https://github.com/ristea/sspcab.

• The paper introduces SSPCAB, a new neural block that integrates self-supervised reconstruction to enhance anomaly detection.
• It employs a masked convolutional layer and channel attention mechanism to boost feature learning and minimize reconstruction error.
• Empirical results on datasets like MVTec AD, Avenue, and ShanghaiTech show significant improvements in key detection metrics.

Overview of the Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection

The paper presents a novel architectural component, the Self-Supervised Predictive Convolutional Attentive Block (SSPCAB), for anomaly detection tasks. Designed to integrate predictive reconstruction capabilities into neural network architectures, SSPCAB aims to improve anomaly detection across various existing image and video frameworks by leveraging self-supervised learning techniques.

Key Components and Methodology

SSPCAB is built upon the concept of reconstructing masked contextual information through a custom-designed neural block. The main elements include:

• Masked Convolutional Layer: A convolutional filter with a masked center is used to learn and predict the missing regions. This component is responsible for capturing local and global features within its receptive field, controlled through a dilation rate, which is adjustable to accommodate specific task requirements.
• Channel Attention Mechanism: SSPCAB employs Squeeze-and-Excitation (SE) networks to enhance or suppress activation channels based on their relevance. This attention mechanism facilitates the balancing of feature maps, reinforcing meaningful representations while attenuating insignificant ones.
• Self-Supervised Reconstruction Loss: A mean squared error (MSE) loss function is utilized to minimize the reconstruction error of the masked regions, which is integrated into the overall training loss of the host architecture.

Results Summary and Analysis

Through empirical experimentation on benchmarking datasets such as MVTec AD, Avenue, and ShanghaiTech, the paper demonstrates the efficacy of SSPCAB when integrated into state-of-the-art anomaly detection frameworks. Notably, SSPCAB facilitates:

• Enhanced performance metrics, showing a significant increase in detection accuracy and precision across image and video anomaly detection tasks, evidenced by increased AUROC, AUC, RBDC, and TBDC scores.
• State-of-the-art performance improvements, specifically on challenging datasets like Avenue and ShanghaiTech, showcasing the block’s capability to derive benefit across different modalities.

Implications and Future Directions

The introduction of SSPCAB has several implications:

• Theoretical Contributions: By embedding a predictive self-supervised task into a neural architecture, SSPCAB not only improves anomaly detection but also provides a new approach to handling missing data within neural networks.
• Practical Applications: The block’s generic nature allows for seamless integration with diverse architectures, potentially benefiting a wide range of applications extending from industrial inspection to public safety and beyond.

Going forward, potential advancements could include adding 3D masking capabilities to address temporal anomalies more effectively or extending applications to other domains requiring nuanced anomaly detection. Moreover, further research is encouraged to refine masked patterns and explore different receptive field designs to optimize kernel performance.

In conclusion, SSPCAB represents a notable advance in the field of anomaly detection, offering a self-supervised enhancement to existing frameworks while contributing a versatile architectural component capable of scalable integration across numerous domains.