Rethinking multi-stage calibration fusion network and loss function for steel surface defect detection.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Surface defects in steel significantly compromise product reliability and lifespan.
• Detecting these defects is challenging due to varying sizes, complex backgrounds, and minute targets, requiring a balance between accuracy and efficiency.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and efficient steel surface defect detection.
• To address the limitations of existing methods in handling diverse defect characteristics and complex industrial environments.

Main Methods:

• Proposed the Multi-stage Calibration Fusion Network (RMCF-Net) incorporating an Efficient Feature Extraction Module (E-C2f) and a Lightweight Feature Aggregation Module (LFAM).
• Introduced a Multi-stage U-shaped Fusion Network (MUF-Net) with adaptive feature re-weighting for enhanced cross-layer feature interaction.
• Developed an adaptive adjustment function (QCIoU) to refine bounding box predictions.

Main Results:

• Achieved state-of-the-art mean Average Precision (mAP@0.5) scores: 78.9% on NEU-DET, 67.1% on GC10-DET, and 71.9% on Track Components.
• Maintained a high detection speed of 119 Frames Per Second (FPS).

Conclusions:

• RMCF-Net demonstrates superior performance in steel surface defect detection, balancing accuracy and efficiency.
• The proposed model offers reliable technical assurance for quality control in intelligent manufacturing.