基于U型网络的K-TIG焊熔池图像分割方法研究
    点此下载全文
引用本文:黄 辉1,蔡庆荣1,陆立明2,李会军1.基于U型网络的K-TIG焊熔池图像分割方法研究[J].计算技术与自动化,2022,(2):100-107
摘要点击次数: 67
全文下载次数: 0
作者单位
黄 辉1,蔡庆荣1,陆立明2,李会军1 (1.五邑大学 广东 江门 5290202.江门云天电力设计咨询有限公司广东 江门 529100) 
中文摘要:小孔钨极惰性气体保护焊K-TIG(Keyhole Tungsten Inert Gas Welding)的强烈氩弧光会对熔池液面与焊件的成像造成干扰。针对其熔池图像呈现边界模糊、特征尺寸差异大、形态不规则的特点,设计了一种基于注意力机制的多尺度特征融合语义分割模型,对焊接熔池区域图像进行分割。首先,利用并行结构的多尺度非对称瓶颈单元替换UNet在特征提取过程中的卷积块,加强对不同尺度特征提取的能力,强化对熔池轮廓的表征能力;然后,在上采样阶段使用跨层的注意力模块引导网络模型更加关注熔池的锁孔区域;最后,在进行网络训练之前,对训练集的图片采用Multi-Scale Retinex算法执行颜色和形态的图像增强。实验结果显示,该神经网络的分割结果与熔池实际区域在Dice系数、MIOU以及F1系数的指标上分别达到95.78%、83.32%和91.86%。
中文关键词:K-TIG焊接  熔池图像分割  编解码结构  注意力机制  多尺度卷积核
 
Research on Image Segmentation Method of K-TIG Welding Pool Based on U-shaped Network
Abstract:The intense argon arc light of keyhole Tungsten inert gas welding (K-TIG) will interfere the image of molten pool surface and welding parts. In view of the fuzzy boundary, large difference of characteristic size and irregular shape of the molten pool image, a multi-scale feature fusion semantic segmentation model based on attention mechanism was designed to segment welding pool image. Firstly, the convolution block of UNet in feature extraction process is replaced by multi-scale asymmetric bottleneck element of parallel structure, which strengthens the ability of feature extraction of different scales and the ability of representation of molten pool contour. Then, in the up-sampling stage, cross-layer attention modules are used to guide the network model to pay more attention to the keyhole region of the molten pool. Finally, multi-scale retinex algorithm was used to perform color and shape image enhancement for the images of the training set before network training. The experimental results show that the segmentation results of the neural network and the actual molten pool area reach 95.78%, 83.32% and 91.86% in Dice coefficient, MIOU coefficient and F1 coefficient respectively.
keywords:K-TIG welding  weld pool image segmentation  codec structure  extrusion and excitation  attention mechanism  multiscale convolution kernel
查看全文   查看/发表评论   下载pdf阅读器