Agm-Net: Attention-guided masking denoising anomaly location network

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised anomaly detection using knowledge distillation (KD) is effective but limited by similar student-teacher network architectures.
• Previous methods struggled with performance and generalization due to architectural constraints.

Purpose of the Study:

• To propose an attention-guided masked denoising anomaly localization network (Agm-Net) to overcome limitations of existing KD-based anomaly detection.
• To enhance structural differences between student and teacher networks and improve anomaly localization accuracy.

Main Methods:

• Incorporated an attention-guided U-shaped denoising architecture into the student network.
• Introduced a feature-level mask generation module with regionally random masking for controllable masked region sizes.
• Developed a randomly connected boundary smoothing anomaly synthesis strategy for realistic defect image generation.

Main Results:

• Agm-Net achieved high performance on benchmark datasets: MVTec AD (98.2% AU-ROC, 94.6% PRO), VisA (99.2% AU-ROC, 95.1% PRO).
• The model demonstrated effectiveness on a real-world PCB dataset (BHAAD) with 98.9% AU-ROC and 93.4% PRO.
• The proposed methods improved feature detail recovery and local information understanding.

Conclusions:

• Agm-Net significantly advances unsupervised anomaly detection by enhancing model architecture and data synthesis.
• The attention-guided denoising and masked feature learning contribute to superior anomaly localization capabilities.
• The model shows strong generalization across diverse datasets, including industrial inspection tasks.