基于改进型U-net神经网络的焊缝图像处理  

Weld Seam Image Processing Based on Improved U-net Neural Network

作　　者：申俊琦[1,2] 刘晨帆 胡绳荪 杨辉[3] 张超 吴定勇 Shen Junqi;Liu Chenfan;Hu Shengsun;Yang Hui;Zhang Chao;Wu Dingyong(School of Materials Science and Engineering,Tianjin University,Tianjin 300354,China;Tianjin Key Laboratory of Advanced Joining Technology,Tianjin 300354,China;Tianhe Mechanical Equipment Manufacturing Co.,Ltd.,Suzhou 215557,China)

机构地区：[1]天津大学材料科学与工程学院,天津300354 [2]天津市现代连接技术重点实验室,天津300354 [3]中交天和机械设备制造有限公司,苏州215557

出　　处：《天津大学学报（自然科学与工程技术版）》2023年第4期436-442,共7页Journal of Tianjin University：Science and Technology

基　　金：国家自然科学基金资助项目(51575381)。

摘　　要：对于基于线结构光的视觉焊缝跟踪系统,焊缝信息提取的精度和速度是焊缝图像处理算法的两个关键指标.对于含有强噪声的焊缝图像,传统的图像处理算法很难达到较高的提取精度;而基于传统全卷积神经网络的图像处理算法则可以有效地提取焊缝信息,但该网络仅对深层的抽象特征进行连续的上采样,忽略了焊缝像素点间空间位置的关系,从而割裂了空间的一致性,降低了焊缝信息提取的精度.针对以上问题,提出基于改进型U-net全卷积神经网络的焊缝图像处理模型以实现焊缝信息的逐级恢复.该模型通过引入双U型结构,将下采样倍率从16倍降为4倍,更多地保留了下采样时的空间信息;通过桥接的方式将第1次下采样时的焊缝特征传入第2次上采样阶段,为抽象的特征信息融入更多的空间信息,提高了焊缝信息提取的精度;将带泄漏的修正线性单元作为神经网络的激活函数,有效避免了原生U-net网络神经元坏死的现象.网络训练结果表明,使用相同数据集训练时,与传统FCN-32s网络和原生U-net网络相比,该模型的像素精度、平均像素精度和平均交集对联合均为最高.实验结果表明:该模型的焊缝位置提取平均偏差为1.64 mm,单帧焊缝图像处理时间为6.4 ms;该模型对含强噪声图像的焊缝信息提取精度和速度均优于Sterger算法和传统FCN-32s网络.For the weld seam tracking system using line structured light vision,the two most important parameters of the image processing algorithm are the accuracy and speed of weld seam information extraction.With conventional image processing algorithms,it is difficult to achieve high precision for weld seam images containing strong interferences.The image processing model using the conventional full convolutional neural network can effectively extract weld seam information;however,only the continuous upsampling of deep abstract features is performed in this type of network,ignoring the spatial relationship between the weld seam pixel positions.Therefore,spatial consistency is divided,and the extraction of weld seam information becomes less precise.To solve these problems,a weld seam image processing model based on an improved U-net full convolutional neural network was proposed to recover weld seam information step by step.A double U-shaped structure was introduced to reduce the downsampling rate from 16 to 4 times,thus retaining more spatial information during the downsampling process.Weld seam information from the first downsampling stage was bridged to the second upsampling stage to integrate more spatial information into abstract weld seam characteristics,thereby improving the accuracy of the weld seam information extraction.To avoid the phenomenon of neuron necrosis in the original U-net,the leaky rectified linear unit was used as the activation function of the neural network.Using the same training conditions as the conventional FCN-32s network and original U-net,the highest pixel accuracy,mean pixel accuracy,and mean intersection over union were obtained.Using the upgraded U-net network,the mean deviation of the weld seam position was 1.64 mm,and the processing time of a single-frame weld seam image was 6.4 ms.This model showed considerable precision and speed improvements over the Sterger algorithm and the conventional FCN-32s network in extracting weld seam information from images containing heavy noise.

关 键 词：U-net神经网络 焊缝图像处理 线结构光视觉 焊缝跟踪 

分 类 号：TG409[金属学及工艺—焊接]