应用神经网络的复杂物体三维测量

Neural Network Applied to Three-Dimensional Measurement of Complex Objects

将神经网络引入基于结构光投影的复杂物体三维面形测量。在测量过程中,利用神经网络强大的函数逼近能力,得到离散条纹图的连续逼近函数,从中解出物体的相位分布信息,获得物体的三维面形分布。应用神经网络方法,在结构光投影条件下,只需要获取一幅条纹图,便可以完成复杂物体的三维面形测量。该方法相比传统的傅里叶变换轮廓术,不存在滤波操作,不会在测量过程中丢失被测物体的高频分量,具有更高的空间带宽积和灵敏度,能准确测量出复杂物体的细节,更加适用于恢复复杂物体的三维面形。并且该方法在条纹图存在阴影的情况下与傅里叶变换轮廓术相比,能更好地提取出物体的相位信息,恢复物体的三维面形。模拟及实验均验证了该方法的可行性。

The neural network has been introduced into the reconstruction of the complex three-dimensional （3D） object based on structured light projection. In the network method, the neural network with powerful function of approximation is used to get the continuous approximate function of the discrete fringe pattern. The measured object can be reconstructed by dealing with the approximate function and drawing phase distribution of the object. As a result, the network method based on structured light projection need only one deformed fringe pattern to reconstruct the tested object. Compared with the Fourier transform profilometry （FTP）, the neural network method without filtering process does not lose high frequency of the measured object. So it has large space bandwidth product and high sensitivity can given out the detail precisely. Therefore, this method performs better than FTP in the threedimensional shape measurement of complex objects. Moreover, compared with FTP, the network method can demodulate more useful phase from the fringe pattern with shadow. Computer simulations and experiment validate the feasibility of this method.