Wavelet-Guided Promotion-Suppression Transformer for Surface-Defect Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Surface-defect detection requires precise pixel-level annotation of irregular defect areas.
• Existing methods often overlook the unique characteristics of defect textures and edges.
• Industrial products exhibit normal surface patterns that differ significantly from defect appearances.

Purpose of the Study:

• To propose a novel framework for accurate surface-defect detection.
• To enhance the network's focus on fine-grained image details.
• To effectively suppress background noise while preserving critical defect features.

Main Methods:

• Introduced a two-stage "promotion-suppression" transformer (PST) framework.
• Employed wavelet features to guide focus on detailed image features.
• Developed a Haar augmentation module for high-frequency detail sensitivity and a quadratic feature-fusion module (QFFM) for noise suppression.

Main Results:

• The Haar augmentation module improved sensitivity to high-frequency details.
• The QFFM effectively suppressed background noise by analyzing noise and defect feature properties.
• The PST framework demonstrated superior performance on DAGM, MT, and CRACK500 datasets.

Conclusions:

• The proposed PST framework accurately locates and classifies surface defects.
• The integration of wavelet features and the QFFM enhances discriminative feature extraction.
• The PST framework offers a significant advancement in industrial surface-defect detection.