WMC-DFINE: An Improved DFINE Model for Aluminum Profile Surface Defect Detection

Area of Science:

• Materials Science
• Computer Vision
• Industrial Engineering

Background:

• Automated inspection of aluminum profiles is crucial for industrial quality control.
• Existing machine vision methods struggle with background texture interference and detecting defects with extreme aspect ratios.

Purpose of the Study:

• To develop an end-to-end defect detection algorithm for aluminum profiles that overcomes current limitations.
• To enhance the accuracy and efficiency of automated surface defect inspection.

Main Methods:

• Introduced the Wavelet-Integrated Frequency Attention (WIFA) module for noise suppression and edge enhancement.
• Designed a Cross-Scale Feature Fusion (CSFF) module to maintain defect feature continuity.
• Incorporated a Multi-Kernel Strip Shuffle (MKSS) module to capture slender scratch features.
• Employed knowledge distillation to create a lightweight, efficient student model.

Main Results:

• WMC-DFINE achieved a mean Average Precision (mAP) of 82.1% on the Tianchi aluminum defect dataset, outperforming YOLOv12 and RT-DETR.
• The distilled WMC-DFINE-distill model showed a 3.2% mAP improvement over the baseline DFINE.
• The distilled model reduced parameters by 47% and achieved 59.75 FPS inference speed.

Conclusions:

• The WMC-DFINE algorithm effectively balances background suppression and defect detail preservation.
• This method offers a practical and efficient solution for real-time industrial defect inspection of aluminum profiles.
• The knowledge distillation strategy yields a highly efficient model for practical deployment.