Abstract: To address the issue of low accuracy in laser streak segmentation of weld seams caused by substantial arc noise in complex welding environments,an enhanced U-Net robust weld seam-recognition algorithm incorporating an attention mechanism was proposed.The ECA-Net was utilized in the feature-fusion process to achieve a weighted fusion of features.Subsequently,a feature-classification structure was supplemented after the encoder structure such that it could output the name of the corresponding type of weld.Since the imbalance between positive and negative samples in the network training will affect the recognition results,Dice-Loss and Focal-Loss were supplemented to the loss function to enhance the robustness and generalization of the model.Additionally,a method of fusing pixel-position and image-type information was proposed to enhance the robustness of weld recognition.The experiments demonstrated that the proposed method obtained favourable experimental results in an environment with interferences such as arc-light,smoke and noise.Therefore,it could satisfy the demand for accuracy and real-time detection that possessed certain application prospects in the actual welding site with arc-light and smoke interferences.

Key words: weld recognition, image segmentation, attention mechanism, U-Net, robustness

摘要： 针对复杂焊接环境下大量弧光噪声造成焊缝激光条纹分割精度低的问题,提出一种融入注意力机制的改进U-Net鲁棒焊缝识别算法。首先,在模型的特征融合过程中使用超强通道注意力机制实现特征的加权融合。然后,在编码器结构之后,加入特征分类结构,使其可以输出焊缝对应类型名称。最后,由于网络训练中正负样本失衡会对识别结果产生影响,在模型的损失函数中添加Dice Loss和Focal Loss来进行修正,以提高模型的鲁棒性和泛化性。另外,在模型训练的过程中提出了一种像素位置信息和图像种类信息融合的方式,以增强焊缝识别的鲁棒性。实验表明,在具有弧光、烟雾噪声等干扰环境下,所提方法得到了较好的实验结果,能够满足检测对精度和实时性的需求,在具有弧光、烟雾等干扰的实际焊接现场中具有一定的应用前景。

关键词: 焊缝识别, 图像分割, 注意力机制, U-Net, 鲁棒性