Surface Defect Detection Based on Adaptive Multi-Scale Feature Fusion.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Quality Assurance

Background:

• Surface defect detection is crucial for industrial quality control.
• Complex backgrounds and diverse defect types challenge existing salient object detection (SOD) methods.
• Accurate defect identification is essential for manufacturing processes.

Purpose of the Study:

• To propose an Adaptive Multi-Scale Feature Fusion Network (AMSFF-Net) for robust surface defect detection.
• To address the limitations of current SOD techniques in industrial settings.
• To enhance the accuracy and reliability of automated quality assurance systems.

Main Methods:

• Developed AMSFF-Net featuring an upsampling fusion module with adaptive weight, global, and differential feature fusion.
• Integrated a spatial attention (SA) mechanism to improve multi-feature map fusion.
• Employed preprocessing techniques including aspect ratio adjustment and random rotation.
• Curated a magnetic tile defect dataset by removing low-quality samples.

Main Results:

• AMSFF-Net demonstrated superior performance in surface defect detection compared to state-of-the-art methods.
• Achieved an S-measure of 0.9038 and an Fβmax of 0.8782.
• Showcased a 1% improvement in Fβmax over existing leading technologies.

Conclusions:

• AMSFF-Net effectively overcomes challenges posed by complex backgrounds and diverse defects in industrial surface inspection.
• The proposed adaptive fusion and spatial attention mechanisms significantly enhance SOD performance.
• The method offers a promising solution for improving automated quality assurance in manufacturing.