改进U-net神经网络的干涉图像齿面域分割方法  

Improved tooth flank region segmentation method of interference image by U-net neural network

作　　者：杨鹏程[1] 张津京 李小成 孟杰 康乐谦 YANG Pengcheng;ZHANG Jinjing;LI Xiaocheng;MENG Jie;KANG Leqian(School of Mechanical and Electrical Engineering,Xi’an Polytechnic University,Xi’an 710048,China)

机构地区：[1]西安工程大学机电工程学院,陕西西安710048

出　　处：《西安工程大学学报》2024年第6期90-97,共8页Journal of Xi’an Polytechnic University

基　　金：陕西省科技厅自然科学基础研究计划-面上项目(2022JM-219)。

摘　　要：激光干涉法是测量齿轮齿面形貌误差的有效方法,利用干涉图像处理技术可以准确获取齿面形貌信息。然而,在实际实验中,散斑噪声和齿面杂散光等现象往往会导致分割精度的降低,从而影响测量的准确性。同时,传统的时域分割方法存在着拟合误差和难以确定合适分割阈值等问题,也会对分割精度造成影响。因此,齿面域分割在干涉图像处理中起着至关重要的作用,可以有效提高测量精确性。为了解决这些问题,提出了一种基于改进的U-net神经网络的齿面域分割方法。首先,在传统的U-net架构中引入SE注意力机制,以提升网络对齿面区域的识别精度。其次,通过收集不同光照环境和入射角度下的齿面物体图像和干涉图像作为训练数据集,建立网络输入与输出之间的映射关系,增强网络的泛化能力。最后,将文中方法与传统分割方法进行对比实验,结果表明:该方法可以有效降低传统方法引入的分割误差,相位跳变修正率约为75.6%,模型的准确度达到了96%,提升了分割精确性和测量准确性,在干涉图像、医学和遥感图像处理等领域有着广泛的应用潜力。Laser interferometry is an effective method to measure the tooth flank topography errors,and the information of tooth flank topography can be obtained accurately by using interferometric image processing technology.However,in practical experiments,speckle noise and tooth flank stray light often lead to the reduction of segmentation accuracy,thus affecting the accuracy of measurement.At the same time,the traditional time-domain segmentation method has the problems such as fitting error and being difficult to determine the appropriate segmentation threshold,which will also affect the segmentation accuracy.Therefore,tooth region segmentation plays an important role in interference image processing,which can effectively improve the measurement accuracy.To solve this problem,this paper presents a tooth flank region segmentation method based on improved U-net neural network.Firstly,the SE attention mechanism was introduced into the traditional U-net architecture to improve the recognition accuracy of the tooth flank region.Secondly,the mapping relationship between the input and output of the network was established to enhance the generalization ability of the network by collecting the tooth flank object images and interference images under different lighting environments and incident angles as the training data set.Finally,an experimental comparison was conducted between the methods presented in the article and traditional segmentation methods.The results show that the proposed method can effectively reduce the segmentation error introduced by the traditional method,the phase jump correction rate is about 75.6%,and the accuracy of the model is 96%,which improves the segmentation accuracy and measurement accuracy.This research has wide application potential in the fields of interference image,medical and remote sensing image processing.

关 键 词：激光干涉测量 干涉图像分割 U-net神经网络 分割精度 

分 类 号：TN247[电子电信—物理电子学]