MSF-Net: Multi-Scale Feature Learning Network for Classification of Surface Defects of Multifarious Sizes

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Surface defect detection faces challenges due to significant variations in defect scale.
• Existing Convolutional Neural Network (CNN) methods struggle with detecting small or local defects, leading to imbalanced feature representation.

Purpose of the Study:

• To propose a novel Multi-Scale Feature Learning Network (MSF-Net) for enhanced surface defect detection.
• To improve the network's ability to capture features across different scales, particularly for small and localized defects.

Main Methods:

• Developed a Dual Module Feature (DMF) extractor using optimized Concatenated Rectified Linear Units (CReLUs) and Inception modules.
• Integrated multi-scale receptive fields by merging feature maps and employed residual connections, batch normalization, and average pooling to optimize training efficiency.

Main Results:

• The proposed MSF-Net demonstrated superior performance in detecting surface defects with multi-scale features.
• Experimental results on two benchmark datasets validated the network's advancement and effectiveness compared to existing methods.

Conclusions:

• MSF-Net offers a more balanced feature expression capability, significantly improving multi-scale surface defect detection.
• The network architecture effectively handles large scale differences in defects, enhancing overall detection accuracy.