基于卷积神经网络UNet++的散斑图像变形测量方法  

Speckle Image Deformation Measurement Method Based on Convolutional Neural Network UNet++

作　　者：陈强 梁浚哲 梁晋[1] CHEN Qiang;LIANG Junzhe;LIANG Jin(State Key Laboratory of Precision Manufacturing Technology,School of Mechanical Engineering,Xi'an Jiaotong University,Xi'an 710049,China)

机构地区：[1]西安交通大学机械工程学院精密制造技术全国重点实验室,西安710049

出　　处：《光子学报》2024年第11期202-213,共12页Acta Photonica Sinica

基　　金：国家重点研发计划(No.2022YFB4601802);国家自然科学基金(Nos.52275543,52305586)。

摘　　要：为了进一步提高深度学习在散斑图像变形测量领域的测量精度和泛化能力,提出了一种散斑图像的变形场测量方法,该方法基于图像分割网络UNet++,并融入残差块和坐标注意力机制分别测量散斑图像的位移场和应变场。为了提高网络的泛化性能,在Hermite数据集的基础上增添了真实试验中的散斑图案以及亮度变化,构建了适应于该网络的全新数据集。对该方法与现有深度学习方法在自建数据集和公开数据集上分别进行测试,结果表明:所提方法在所有测试中均取得了最高的平均精度和最优的鲁棒性;其他网络在具有亮度变化的数据集中几乎失效,而所提网络依然能准确测量出变形场;在DIC挑战数据集中Star5图像集和Star6图像集上,所提网络获得了最低的性能度量指标,分别为1.372和0.003 7。Traditional Digital Image Correlation(DIC)methods face challenges in terms of computational speed,especially for large datasets,and in handling complex scenarios involving high-frequency deformation or discontinuities such as cracks.With the advent of deep learning,the potential for leveraging Convolutional Neural Networks(CNNs)for DIC has become increasingly apparent.Deep learning has revolutionized computer vision,achieving state-of-the-art results in tasks such as image classification,object detection,and segmentation.The success of CNNs in these domains suggests that they could also be applied to the task of DIC,potentially offering improvements in both accuracy and computational efficiency.In this context,the publication presents a novel approach to DIC using an advanced CNN architecture known as UNet++.The proposed method,termed DIC-Net++,is designed to address the limitations of traditional DIC algorithms and enhance the performance of deep learning in the context of speckle image deformation measurement.The article has developed two specialized networks within the DIC-Net++framework:DIC-Net++-d for displacement field measurement and DIC-Net++-s for strain field measurement.These networks are built upon the UNet++architecture,which is known for its effective feature extraction and fusion capabilities,and have been augmented with residual blocks and coordinate attention mechanisms to improve their performance.To facilitate the training of these networks and enhance their generalization capabilities,a new dataset is constructed that extends the Hermite dataset with additional real experimental speckle patterns and variations in brightness.This comprehensive dataset includes 47800 image pairs,with 35800 pairs derived from the Hermite dataset,10000 from real experiments,and 2000 involving high-frequency deformation.The dataset is meticulously divided into training,validation,and testing subsets to ensure a robust evaluation of the proposed networks.The text describes the design and training process of the DIC-

关 键 词：光学测量 形变测量 卷积神经网络 数字图像相关法 深度学习 

分 类 号：O348.1[理学—固体力学]