DAMF-Net: A Dynamic Receptive Field Enhancement and Semantic-Guided Adaptive Modulated Fusion Network for Steel...

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Industrial steel production faces challenges in defect segmentation due to complex backgrounds and varying defect scales.
• Existing semantic segmentation models struggle with scale confusion and noise propagation, limiting accuracy and stability.

Purpose of the Study:

• To develop an advanced steel surface defect segmentation network (DAMF-Net) addressing limitations of current models.
• To enhance segmentation accuracy, stability, and efficiency in complex industrial environments.

Main Methods:

• Introduced a Kernel Selection Fusion Attention (KSFA) module for enhanced multi-scale feature representation in the encoding stage.
• Developed an Adaptive Modulated Fusion Module (AMFM) to reduce noise propagation by adaptively fusing shallow and deep features in the decoding stage.
• Utilized a hybrid optimization objective (binary cross-entropy and Dice loss) to improve detection of small-scale and low-contrast defects.

Main Results:

• DAMF-Net achieved significant improvements in mDice/Dice metrics (2.38% and 4.40%) over the U-Net baseline on Severstal and ESDIs-SOD datasets.
• Demonstrated lower prediction errors and more stable segmentation performance under complex background conditions.
• Achieved a high inference speed of 55.8 FPS, balancing precision and efficiency.

Conclusions:

• DAMF-Net offers an effective solution for high-precision steel surface defect segmentation in industrial scenarios.
• The proposed dynamic receptive field modulation and adaptive fusion techniques significantly enhance segmentation performance.
• The network provides a robust and efficient tool for quality control in steel manufacturing.