多方法相融合的复杂物体深度信息的恢复

Recovering the Structure of Complex Objects by an Integrated Method

在恢复图象深度信息的方法之中，利用立体视觉的偏差来精确地定位物体的深度，是行之有效的，但只能适用于可匹配的特征点，如何建立左右图象中对应点的匹配是该方法的主要障碍；ＳｈａｐｅＦｒｏｍＳｈａｄｉｎｇ方法是利用单幅图象的灰度信息获取物体表面的形状信息（表面的方向），而不能获得其深度信息，其约束条件是表面的光滑性。在此用神经网络方法将二者融合起来，形成优势互补，用来获取物体的深度信息，通过对合成图象及实际图象进行的实验。

A new method is presented for recovering the 3 dimensional structure of complex objects. Stereo is an effective method to determine the precise depth of the surface through binocular disparity. But it is difficult to establish the correspondence of the two image. The method of Shape From Shading (SFS) can learn the surface shape but not the surface depth of an object from merely one image under the constraint of surface smoothness; In this paper, the methods of stereo and SFS are integrated into neural networks to locate the precise depth and shape information of the object, which gives full play to their superiority. Experimental results with numerically generated and laboratory images are given to verify the method.