An efficient triple-domain YOLOv8 for real-time aluminum profile defect detection

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Aluminum surface defect detection faces challenges from complex textures, illumination variations, and reflective noise.
• Conventional methods using spatial-domain features struggle with subtle, texture-oriented, and elongated defects.

Purpose of the Study:

• To develop a real-time framework for improved aluminum surface defect detection.
• To enhance detection accuracy by integrating multi-domain features.

Main Methods:

• Proposed Adaptive Triple-Domain YOLOv8 framework integrating spatial, frequency (Discrete Wavelet Transform), and gradient (Gabor filter) features.
• Utilized an adaptive attention-based fusion module for dynamic multi-domain feature integration.
• Evaluated on the Tianchi benchmark dataset.

Main Results:

• Achieved a mean Average Precision (mAP@0.5) of 96.7% and mAP@0.5:0.95 of 73.2%.
• Demonstrated superior performance over spatial-domain, dual-domain, and recent detectors.
• Maintained real-time inference speed exceeding 100 Frames Per Second (FPS).

Conclusions:

• The proposed method significantly improves the detection of low-contrast, texture-dominated, and elongated defects.
• Offers a balanced solution for industrial aluminum surface inspection, prioritizing both accuracy and real-time efficiency.