SD-GASNet: Efficient Dual-Domain Multi-Scale Fusion Network with Self-Distillation for Surface Defect Detection

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Surface defect detection is crucial for industrial quality control.
• Deep learning models struggle with subtle defects, large variations, and limited computational resources.
• Existing methods lack accuracy and real-time performance in diverse industrial settings.

Purpose of the Study:

• To develop an efficient and accurate deep learning network for industrial surface defect detection.
• To address challenges of subtle defects, scale variations, and computational constraints.
• To improve generalization across different sensor types and datasets.

Main Methods:

• Proposed SD-GASNet, a network employing a self-distillation model compression strategy.
• Introduced Alignment, Enhancement, and Synchronization Feature Pyramid Network (AES-FPN) with Frequency Domain Information Gathering-and-Allocation (FIGA) and Channel Synchronization (CS) modules.
• Utilized Multi-scale Feature Alignment (MFA) and Frequency-Guided Perception Enhancement Module (FGPEM) for feature refinement and enhancement, alongside Enhanced KL divergence loss for self-distillation.

Main Results:

• SD-GASNet achieved state-of-the-art performance on three public datasets (NEU-DET, PCB, TILDA).
• Demonstrated excellent generalization capabilities across different industrial imaging applications.
• Achieved superior accuracy and a competitive inference speed of 180 FPS.

Conclusions:

• SD-GASNet offers a robust and generalizable solution for sensor-based industrial imaging.
• The proposed methods effectively address challenges in detecting subtle defects and resource constraints.
• The network provides high-precision, real-time surface defect detection vital for industrial quality control.