Multiscale diffusion-enhanced attention network for steel surface defect detection in Polysilicon Production

Area of Science:

• Materials Science
• Computer Vision
• Industrial Automation

Background:

• Surface defect detection is critical for steel component quality control in polysilicon production.
• Challenges include tiny defects, irregular shapes, complex backgrounds, and low contrast, hindering traditional methods.

Purpose of the Study:

• To develop an advanced deep learning network for accurate and efficient surface defect detection on steel components.
• To address the limitations of existing methods in detecting small, irregular, and low-contrast defects.

Main Methods:

• Proposed MSEOD-DDFusionNet (Multi-Scale and Effective Object-Detection Diffusion Fusion Network), a multi-scale diffusion-enhanced attention network.
• Integrated specialized modules: MTECAAttention, ODConv, LMDP, and DDFusion.
• Applied network pruning to optimize computational complexity and enhance accuracy.

Main Results:

• Achieved state-of-the-art performance on the specialized DDTE dataset (82.6% mAP, 61.6% F1-score).
• Demonstrated high inference speed (193.5 FPS) with a small parameter count (8.46M).
• Showcased excellent generalization capabilities across public benchmarks (NEU-DET, GC10-DET) and cross-domain tasks.

Conclusions:

• MSEOD-DDFusionNet provides an efficient and accurate solution for industrial surface defect inspection.
• The proposed network effectively handles challenges like tiny defects and complex backgrounds.
• The model's strong performance and generalization pave the way for improved quality control in steel manufacturing.