Steel Surface Defect Detection Based on Improved RT-DETR
Qiubo HUANG , Yakui ZHANG , Mei WANG
Journal of Donghua University(English Edition) ›› 2026, Vol. 43 ›› Issue (5) : 80 -89.
To address the challenges of low accuracy in small target detection and high model complexity in steel surface defect detection, a novel defect detection algorithm named FCS-DETR is proposed based on the real-time detection transformer (RT-DETR) framework. The methodology involves three key technical improvements. First, a newly designed DC-RepC3 module replaces the original RepC3 module to expand the model’s receptive field for enhanced contextual information capture, which effectively suppresses complex background noise while improving detection accuracy. Second, for selective feature extraction, partial convolution is integrated into the backbone network to significantly reduce the parameter count and computational load, while wavelet pooling is further incorporated to strengthen model robustness. Finally, a Wise-ShapeIoU-NWD joint loss function is proposed, which incorporates both the shape and distribution characteristics of the target, reducing the geometric discrepancies between the predicted and ground truth bounding boxes and thus improving bounding box regression accuracy. Experimental results on the NEU-DET dataset demonstrate that FCS-DETR achieves a mean average precision at an intersection-over-union threshold of 0.50 (mAP50) of 76.6%, representing an improvement of 3.4 percentage points over RT-DETR-R18, while reducing the parameter count and computational complexity of the model by 31.8% and 28.8%, respectively. These advancements establish FCS-DETR as an effective solution for accurate and efficient steel surface defect detection, showing substantial potential for industrial applications.
steel surface defect detection / RT-DETR / partial convolution / dilated convolution / wavelet pooling / Wise-IoU
| [1] |
|
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
Zhang H, |
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
|
| [19] |
Huang H, Chen Z, Zou Y, |
/
| 〈 |
|
〉 |