SMF-DETR: An Efficient Lightweight Detection Transformer for Real-Time Bearing Surface Defect Detection

Area of Science:

• Mechanical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Industrial equipment reliability depends on effective bearing surface defect detection.
• Current deep learning methods face challenges with small targets, high computational load, and edge deployment.
• There is a need for efficient and accurate defect detection algorithms suitable for industrial settings.

Purpose of the Study:

• To propose an efficient defect detection algorithm, SMF-DETR, addressing limitations of existing methods.
• To enhance accuracy for small bearing surface defects.
• To reduce computational complexity and enable edge device deployment for real-time applications.

Main Methods:

• Developed the StarNet-MEIS-FDConv-detection transformer (SMF-DETR) algorithm.
• Incorporated element-level multiplication for high-dimensional feature mapping and reduced complexity.
• Utilized a multiscale edge information selection mechanism for improved small defect detection.
• Employed frequency domain dynamic convolution for efficient and adaptive feature extraction.

Main Results:

• Achieved 96.2% mAP@50 and 98.1% accuracy on custom bearing defect datasets.
• Outperformed baseline methods by 3.1% in mAP@50 and 2.9% in accuracy.
• Reduced computational cost by 57.7% and model size by 37.1%.
• Demonstrated real-time processing speeds (97.3 FPS on desktops, 58.1 FPS on RK3588).

Conclusions:

• SMF-DETR offers a significant improvement in bearing surface defect detection accuracy and efficiency.
• The algorithm is suitable for real-time industrial applications, including on embedded platforms.
• Validation on public datasets confirms the algorithm's versatility and generalization capabilities.