基于深度学习方法的单张零件图像重建网格模型  

Reconstructing mesh model from single part images based on deep learning method

作　　者：田铮 龙雨 TIAN Zheng;LONG Yu(State Key Laboratory of Featured Metal Materials and Life-cycle Safety for Composite Structures,Guangxi University,Nanning 530004,China)

机构地区：[1]广西大学省部共建特色金属材料与组合结构全寿命安全国家重点实验室,广西南宁530004

出　　处：《现代电子技术》2024年第9期109-114,共6页Modern Electronics Technique

基　　金：广西自然科学基金项目(2023GXNSFBA026287);广西重点研发计划(桂科AB23026101)。

摘　　要：重建物体的三维形状是计算机图形学领域的一个研究热点,网格模型是一种常用的三维模型。文中基于深度学习方法提出一种从单幅机械零件图像重建其网格模型的方法。首先经过一个图像预处理过程将前景零件从背景中分离出来;其次基于ResNet和BSP⁃Net两个骨干网络建立一个新的网络结构,将前景零件图像重构为网格模型。该网络将零件的多个视图图像作为输入,并融合它们的重要特征。此外,加入形状先验损失引导模型的训练过程以优化重建结果。对螺母、螺栓和垫圈进行重建实验,验证了该方法的有效性。通过训练过程的损失函数曲线说明添加多视图特征融合和形状先验损失可以让损失收敛到更低的值。在三个评价指标上的测试表明,文中方法的重建结果优于ResNet+BSP⁃Net方法和IM⁃Net方法。Reconstructing the 3D shape of an object is a research hotspot in the field of computer graphics.Mesh model is a commonly used 3D model.In this paper,a method to reconstruct a mesh model from a single image of a mechanical part is proposed based on deep learning method.The foreground parts are separated from the background after a process of image preprocessing.A new network structure is built based on two backbone networks(ResNet and BSP⁃Net)to reconstruct the foreground part images into a mesh model.In this network,multi⁃view images of the parts are taken as the input and their important features are fused.In addition,a shape prior loss is added to guide the training process of the model to optimize the reconstruction results.Reconstruction experiments on nuts,bolts and washers validate the effectiveness of the proposed method.The loss function curve of the training process shows that adding multi⁃view feature fusion and shape prior loss can make the loss converge to a lower value.Tests on three evaluation indexes show that the reconstruction results of the proposed method outperform those of the ResNet+BSP⁃Net method and the IM⁃Net method.

关 键 词：网格模型 三维重建 深度学习 特征融合 形状先验 图像分割 

分 类 号：TN911-34[电子电信—通信与信息系统] TP301[电子电信—信息与通信工程]