YOLOv11-WBD: A wavelet-bidirectional network with dilated perception for robust metal surface defect detection.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Metal surface defect detection is crucial for quality control but challenging due to noise, spectral aliasing, and small, low-contrast defects.
• Existing object detection models like YOLO struggle with complex industrial image conditions.

Purpose of the Study:

• To enhance the feature representation and semantic mining capabilities of YOLO for metal surface defect detection.
• To develop a robust and accurate automatic defect detection system for industrial applications.

Main Methods:

• Proposed YOLOv11-WBD model integrating wavelet decomposition, cross-attention, and U-shaped dilated convolution.
• Introduced Wavelet-Attentive Multiband Fusion Module (WAMF) for adaptive multi-frequency feature fusion.
• Developed Bottleneck-Enhanced Dilated U-Conv Module (BEDU) for efficient multi-scale feature aggregation.
• Implemented Bidirectional Depthwise Cross-Attention Module (BDCA) for adaptive feature fusion.

Main Results:

• YOLOv11-WBD achieved performance gains on NEU-DET and GC10-DET datasets, with mAP@0.5 increasing by 5.8% and 2.8% respectively.
• The model demonstrated significantly improved noise tolerance, maintaining high detection accuracy in noisy environments.
• Experimental results validate the effectiveness of the proposed modules in enhancing defect detection.

Conclusions:

• YOLOv11-WBD offers a valuable solution for industrial metal surface defect detection, outperforming existing methods.
• The integration of wavelet decomposition, cross-attention, and dilated convolutions enhances model robustness and accuracy.
• The developed modules provide effective strategies for handling challenging industrial imaging conditions.