基于全特征融合的小径管焊接缺陷检测方法

Welding defect detection method based on full feature fusion for small diameter pipes

针对小径管焊接X射线图像中缺陷尺寸差异大、大纵横比缺陷和回归位置偏移等问题,提出了基于全特征融合和多级检测头的检测模型,模型由全特征融合网络(FFF-Net)和基于距离交并比(IoU)散度(DI-KL)损失的多级检测头组成。FFF-Net通过双向特征均衡有效提取不同尺寸缺陷的特征,利用形状特征提取大纵横比缺陷的形状特征,并通过注意力机制提高特征的显著性;基于DI-KL损失的多级检测头将DI-KL损失作为回归损失,结合多级检测头缓解了预测框回归位置不准确的问题。实验结果表明:该模型有效提高了小径管焊接缺陷检测精度,特别是小尺寸和大纵横比缺陷的检测精度。

Aiming at the problems that size of defects has large difference,large aspect ratio defects,and regression position offset in welding X-ray images of small diameter pipes,a detection model based on full feature fusion and multi-level detection head is proposed,which consists of full feature fusion network(FFF-Net)and multi-level detection head based on distance intersection over union(IoU)Kullback-Leibler divergence(DI-KL)loss.FFF-Net can effectively extract defects features of different sizes by two-way feature equalization.Shape feature can be used to extract the shape features of defects with large aspect ratios,the saliency of features can be improved by attention mechanism,the multi-level detection head based on DI-KL loss is taken as the regression loss,which alleviates inaccurate regression position of the prediction box combing with multi-level detection head.Experimental results show that the model effectively improves the detection precision,especially the detection precision of defects with small size and defects with large aspect ratio.