摘要
为提高焊接件缺陷检测流程的准确性和效率,减少漏检与误检情况,提出了一种基于改进RTDETR的焊接件缺陷检测方法。该方法在RT-DETR模型基础上进行了优化:一方面引入ADown下采样模块,对特征图尺寸进行压缩,同时保留关键信息;另一方面引入WTConv卷积模块,融合多尺度的频域与空间特征,从而增强模型对多尺度目标和复杂背景的特征提取能力。实验结果表明,加入这2个模块后,模型在mAP@50指标上由原RT-DETR的67.6%提升至70.1%,同时参数量由19885212降至18623772。在降低计算复杂度的同时,检测精度得到提升,显著提高了焊接件缺陷检测的效率。
A welding defect detection method based on improved RT-DETR is proposed to improve the accuracy and efficiency of the welding defect detection process,reduce missed and false detections.This method has been optimized based on the RT-DETR model:on the one hand,the ADown downsampling module is introduced to compress the feature map size while preserving key information;On the other hand,the WTConv convolution module is introduced to integrate multi-scale frequency domain and spatial features,thereby enhancing the model̓s feature extraction ability for multi-scale targets and complex backgrounds.The experimental results show that after adding these two modules,the model mAP@50 indicator has increased from 67.6%of the original RT-DETR to 70.1%,while the parameter count has decreased from 19885212 to 18623772.While reducing computational complexity,the detection accuracy has been improved,significantly enhancing the efficiency of defect detection in welded components.
作者
刘鑫
LIU Xin(School of Computer Science,Yangtze University,Jingzhou,Hubei 434023,China)
