Improving Unsupervised Ultrasonic Image Anomaly Detection via Frequency-Spatial Feature Filtering and Gaussian...

Area of Science:

• Non-destructive testing (NDT)
• Artificial Intelligence (AI)
• Computer Vision

Background:

• Ultrasonic image anomaly detection is hindered by limited labeled data, noise, and diverse defect types.
• Existing methods struggle with the complexities of real-world chip packaging defects.

Purpose of the Study:

• To introduce UltraChip, a large-scale C-scan benchmark dataset for ultrasonic anomaly detection.
• To present FSGM-Net, a fully unsupervised framework for accurate, pixel-level anomaly detection.
• To advance annotation-free ultrasonic non-destructive testing (NDT) for practical applications.

Main Methods:

• Developed UltraChip dataset with ~8,000 real-world C-scan images and pixel-level annotations.
• Proposed FSGM-Net, an unsupervised framework utilizing adaptive Frequency-Spatial filtering and an Adaptive Gaussian Mixture Model (Ada-GMM).
• Implemented novel filter loss and entropy-based sparse gating for improved feature consistency and normality weighting.

Main Results:

• FSGM-Net achieved state-of-the-art performance on the UltraChip benchmark.
• Demonstrated superior cross-domain generalization capabilities on MVTec-AD and VisA datasets.
• Achieved real-time inference speeds on a single GPU.

Conclusions:

• The UltraChip dataset and FSGM-Net framework significantly improve ultrasonic anomaly detection capabilities.
• The proposed methods offer robust, annotation-free solutions for practical NDT applications.
• FSGM-Net's effectiveness and efficiency pave the way for wider adoption in industrial inspection.