【摘要】 单频条纹截断相位图的展开是条纹结构光视觉中的一个难题。传统的可靠度引导算法和最小范数算法,在展开有低对比度、条纹断裂、图像离焦或运动模糊的单频条纹的截断相位图时,仍面临很大的挑战。针对这些问题,本文提出一种采用深度卷积神经网络的新方法。截断相位图的展开被分为不可靠区域检测和可靠区域展开两个子问题,深度卷积神经网络被用于检测截断相位图中的不可靠区域,可靠区域的相位展开用传统算法完成。为训练能处理高分辨率截断相位图的神经网络,本文提出位移采样方法。实验采用真实牙模三维扫描数据对神经网络进行训练,并和传统相位展开算法进行对比,验证本文方法的有效性。更多还原

【Abstract】 Phase unwrapping of single frequency wrapped phase map is a difficult problem in fringe structured light vision. Traditional reliability guided algorithms and minimum norm algorithms still face great challenges when unwrapping wrapped phase maps are generated from single frequency fringe images with low contrast, fringe discontinuity, image defocusing, and motion blur. In view of these problems, a new phase unwrapping method using deep convolutional neural networks is proposed. The unwrapping process is divided into two sub-problems: unreliable areas detection and reliability area unwrapping. Deep convolutional neural network is used to detect the unreliable areas of a phase map, and the reliable parts of a phase map is unwrapped with traditional algorithms. In order to train a neural network capable of processing high resolution phase maps, the paper employs a shift-sampling method. Experiment results based on real dental model threedimensional scanning data verify the effectiveness of the prosed method.更多还原