基于神经网络的由单幅二维工程图形重建三维形体系统的研究

3D Objects Reconstruction from Single 2D Graph Based on Neural Networks

摘要: 针对基于单幅图形的三维重建方法的多解性和病态性的难点问题,设计实现了一套三维重建系统,提出一种基于角点特征和图形结构特征的人工神经网络分类算法,对输入图形进行识别分类,然后分别对识别结果精确重建,从而避免了直接恢复深度信息的病态解问题.为有效地提取角点特征,采用改进的基于曲率的自适应角点检测算法,给出了滑动窗口的自适应调整策略,并得到图像边界点的局部支持域的特征向量,从而使提取的角点曲率特征具有旋转、平移和尺度不变性.实验验证了改进算法的快速、准确和稳定的特性.

Abstract: A system scheme was presented in this paper, which could be used in 3D objects reconstruction from single 2D sketch.Corner points and several other structural features were adopted in this system.BP neural network was also applied due to its non-linear mapping and feature extracting ability.In order to obtain the corner points more effectively, an adaptive corner detection algorithm was proposed based on curvature and slide windows.Experimental results show that the selected features and the algorithm have superiority in 3D objects reconstruction.