Surface defect detection of magnetic tile based on RT-DETR improved algorithm

Advances in science and technology have led to an increased use of electric motors in various applications. The quality of magnetic tiles has a direct impact on the motor’s service life, underscoring the importance of detecting defects on the surface of these tiles. In order to address the challenges posed by the small dimensions, diverse shapes, and diminished contrast of magnetic tile defects, which result in suboptimal detection accuracy, frequent missed detections, and diminished efficiency of conventional detection methods, this paper proposes an enhanced algorithm based on RT-DETR and designs a magnetic tile surface defect detection system. In this enhanced algorithm, the conventional upsampling method is substituted by the Dysample in order to enhance the model’s capacity for retaining detail and its anti-aliasing capabilities, thus leading to an improvement in the accuracy of defect detection for small targets. Additionally, a DSBlock module, based on the DSConv, is proposed for integration into RepC3 with a view to enhancing the model’s ability to capture features at multiple scales. Furthermore, the RT-DETR loss function is substituted with WIoUv3 to address issues such as defect overlap, which can hinder the efficacy of defect detection. The experimental findings demonstrate that the AP, AP₅₀, and AR${}_{max=100}$ of the enhanced RT-DETR-DRW model are improved by 1.0 %, 1.6 %, and 1.7 %, respectively, in comparison with the baseline model. RT-DETR-DRW has been shown to outperform models such as YOLOv8l, RT-DETR-x and DFine-s in terms of detection accuracy and leakage for defect detection.