一种用于立体匹配的改进的神经网络方法

An improved Neural Network Utilized in Stereo Matching

实现了一种用于静态视觉立体匹配的神经网络方法。文献［１］提出了一种用于静态体视匹配的神经网络方案，其方案用于随机点图对时存在严重的缺陷。针对随机点图对的特点，对神经网络的偏置输入进行了修正，改进的神经网络能够有效地提取随机点图对中的立体深度信息。为了进一步提高收敛速度和平滑边缘特征，又在偏置输入中引入射线特征，改进了神经元的初始化。在工作站上进行的大量实验模拟表明，我们所做的改进提高了网络的迭代速度和视差图的边缘特征平滑性。

An improved neural network utilized in stereo matching is presented in this paper. The theory for binocular vision and the structure of a recurrent neural network are first described, then some improvements to the neural network are given, including the bias inputs and the initializations of the neurons. Finally, computer simulation is performed. The simulated result supports our improvements.