| 摘 要: 为解决钢材表面缺陷检测中漏检和误检问题,提出一种改进RT-DETR的钢材表面缺陷检测算法。首先,使用双域边缘增强模块DDEEM(Dual Domain Edge Enhancement Module)在时域和频域提取特征;其次,采用高效多尺度注意力机制(Efficient Multi-Scale Attention)捕捉多尺度空间信息;最后,使用基于辅助边框的损失函数(Inner_IoU)更准确地评估边界框重叠区域。实验结果显示,在GC10-DET和NEU-DET数据集上,相较于原始RT-DETR-r18,改进算法的平均精度均值mAP@0.5分别提升2.2个百分点和3.9个百分点。以上结果验证了所提改进策略的有效性,为未来钢铁产业的高度智能化提供了技术支撑。 |
| 关键词: 缺陷检测 深度学习 特征融合 损失函数 RT-DETR |
|
中图分类号:
文献标识码: A
|
| 基金项目: 中共甘肃省委宣传部2023年农村电影公益放映第三方监测服务项目(2023zfcg00315) |
|
| Steel Surface Defect Detection Algorithm Based on Improved RT-DETR |
|
JIA Jianhui1, XIANG Zhong1,2
|
(1.School of Information Science and Engineering, Zhejiang Sc-i Tech University, Hangzhou 310000, China;
2.School of Mechanical Engineering, Zhejiang Sc-i Tech University, Hangzhou 310000, China)
jh77cf77@163.com; xz@zstu.edu.cn
|
| Abstract: To address the issues of missed and false detections in steel surface defect detection, this paper proposes an improved RT-DETR-based steel surface defect detection algorithm. First, a Dual Domain Edge Enhancement Module(DDEEM)is used to extract features in both the time and frequency domains. Second, an Efficient Mult-i Scale Attention mechanism is employed to capture mult-i scale spatial information. Finally, an auxiliary bounding box-based loss function (Inner _ IoU) is utilized to more accurately evaluate the overlapping regions of bounding boxes. Experimental results show that on the GC10-DET and NEU-DET datasets, compared to the original RT-DETR-r18,the improved algorithm achieves increases of 2. 2 percentage points and 3. 9 percentage points, respectively, in mean Average Precision (mAP @ 0.5). These results demonstrate the effectiveness of the proposed improvement strategies and provide technical support for the future high-level intelligence of the steel industry |
| Keywords: defect detection deep learning feature fusion loss function RT-DETR |
