MCM-YOLO: A Lightweight Conflict Mitigation Network for Industrial Metal Surface Defect Detection

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Industrial metal surface defect detection is crucial for quality control.
• Lightweight models are needed for efficient deployment but struggle with defect perception.
• Challenges include weak textures, blurred boundaries, and small defect scales.

Purpose of the Study:

• To propose a lightweight You Only Look Once (YOLO)-based framework, MCM-YOLO, for industrial metal surface defect detection.
• To enhance fine-grained local representation and multiscale feature aggregation.
• To address Feature Integration Conflict (FIC) between enhanced backbone and neck modules.

Main Methods:

• Developed MCM-YOLO based on YOLOv11.
• Introduced C3-Res2Lite (C3-R2L) block for enhanced local representation.
• Implemented Aggregated Bidirectional Feature Pyramid Network (Agg-BiFPN) for feature aggregation.
• Designed a synergistic attention block (MCM-SAB) to mitigate FIC.

Main Results:

• MCM-YOLO achieved 80.4% mAP on NEU-DET and 65.1% mAP on GC-10 (IoU=0.5).
• The model has low computational cost: 6.3 GFLOPs and 2.70 million parameters.
• Demonstrated a competitive accuracy-efficiency tradeoff.

Conclusions:

• MCM-YOLO effectively balances defect detection accuracy and computational efficiency.
• The proposed framework is suitable for industrial visual inspection under resource constraints.
• MCM-SAB successfully alleviates feature incompatibility issues.