Localize-diffusion based dual-branch anomaly detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Limited real anomaly samples hinder effective anomaly detection model training.
• Existing data augmentation methods generate unrealistic and non-diverse pseudo anomalies.
• Current methods often place augmented anomalies in inappropriate locations, reducing practical value.

Purpose of the Study:

• To develop a novel data augmentation technique for generating diverse and realistic anomaly samples.
• To propose a dual-branch anomaly detection (DBA) method that leverages these enhanced samples.
• To improve the accuracy and efficiency of anomaly detection systems.

Main Methods:

• Introduced Localize-Diffusion (LD) augmentation to infer anomaly position and size based on color distribution.
• Implemented hard augmentation and patch propagation to enrich sample diversity.
• Developed a dual-branch network focusing on anomaly-specific and residual features.
• Utilized an adaptive scoring module for weighted fusion of branch outputs.

Main Results:

• The proposed Localize-Diffusion (LD) augmentation effectively generates realistic and diverse pseudo anomaly samples.
• The dual-branch anomaly detection (DBA) method achieved high performance with only 14.2M parameters.
• Achieved a detection Area Under the Curve (AUC) of 99.6 on the MVTec AD dataset.

Conclusions:

• The proposed DBA technique with LD augmentation offers a powerful solution for anomaly detection with limited data.
• DBA demonstrates superior performance and efficiency compared to existing methods.
• The approach effectively addresses the limitations of traditional data augmentation in anomaly detection.