基于卷积神经网络的立体匹配算法

Stereo matching based on convolution neural network

针对传统立体匹配算法特征提取能力有限,多步骤、多模块的任务处理方式容易造成误差累积并且计算过程繁冗复杂等问题,本文在基于卷积神经网络立体匹配算法基础上改进了网络结构,设计了一个端到端的六层卷积神经网络框架预测视差,可以生成保留细节信息良好的视差图。

In view of the limited feature extraction ability of the traditional stereo matching algorithm, the multi-step and multi module task processing method is easy to cause error accumulation and the calculation process is complex, this paper improves the network structure based on the convolution neural network stereo matching algorithm, designs an end-to-end six layer convolution neural network framework to predict disparity, which can generate dispirty map with good details information.