HCFNet: A SAM2-Based Hierarchical Cross-Branch Frequency-Aware Network for Industrial Surface Defect Segmentation

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Foundation models like Segment Anything Model 2 (SAM2) excel at image segmentation but struggle with industrial defect detection due to domain gaps and limited sensitivity to fine details.
• Existing methods often require manual prompts, hindering seamless application in industrial settings.

Purpose of the Study:

• To adapt SAM2 for prompt-free industrial defect segmentation, addressing the domain gap and improving sensitivity to subtle defects.
• To develop a novel network architecture that leverages pre-trained foundation models for automated industrial inspection.

Main Methods:

• Proposes the Hierarchical Cross-Branch Frequency-Aware Network (HCFNet) integrating a Gated Adapter for efficient cross-domain transfer with a frozen SAM2 encoder.
• Introduces a Laplacian-enhanced Auxiliary Branch to amplify high-frequency components, improving detection of fine-grained defects like micro-cracks.
• Employs a Cross-branch Multi-scale Fusion Module to integrate global semantic information with local structural details.

Main Results:

• HCFNet consistently outperforms SAM2-based baselines on MVTec AD and VisA datasets in terms of mean Intersection over Union (mIoU) and mean Dice coefficient (mDice).
• The proposed method demonstrates enhanced sensitivity to subtle industrial defects.

Conclusions:

• HCFNet provides an effective approach for adapting foundation models for prompt-free industrial defect segmentation.
• This work advances precise defect perception technologies for automated industrial inspection.