ABSTRACT Bearing surface defect detection is critical for industrial equipment reliability, but existing deep learning methods suffer from low accuracy for small targets, high computational complexity, and limited edge device deployment. This paper proposes an efficient defect detection algorithm based on the StarNet‐MEIS‐FDConv‐detection transformer (SMF‐DETR). The algorithm employs element‐level multiplication operations in the backbone network to achieve high‐dimensional feature mapping, effectively reducing computational complexity while improving feature extraction capability. The multiscale edge information selection mechanism processes features at different resolutions simultaneously to improve small defect detection. Frequency domain dynamic convolution adapts to different frequency components for optimal feature extraction while maintaining computational efficiency. Experiments on custom bearing defect datasets show that SMF‐DETR achieves 96.2% mean average precision@50 (mAP@50) and 98.1% accuracy, improving baseline performance by 3.1% and 2.9%, respectively. The model also reduces computational cost by 57.7% and model size by 37.1%. Processing speeds reach 97.3 frames per second (FPS) on desktop systems and 58.1 FPS on embedded RK3588 platforms, meeting industrial real‐time detection requirements. Finally, experimental validation was conducted on the publicly available bearing defect‐detection dataset and the PASCAL visual object classes dataset, demonstrating the algorithm's versatility and generalization capabilities.
Gao et al. (Sun,) studied this question.