一种基于卷积神经网络的立体匹配算法设计

A stereo matching algorithm based on convolutional neural network

为了解决传统立体匹配算法对立体图像在低纹理以及遮挡区域匹配效果较差的问题,设计了一种端到端的基于卷积神经网络(Convolutional Neural Network,CNN)的立体匹配算法。该算法采取了残差卷积神经网络对图像特征进行提取,之后利用空洞空间金字塔池化(Atrous Spatial Pyramid Pooling,ASPP)模块来获取图像的上下文信息,并结合多尺度的三维卷积神经网络对代价空间进行规整,最终实现了高精度的立体匹配算法。所获取的视差图在KITTI2015测试平台上的误匹配率为2.42%,与几何上下文(Geometry and Context,GC)网络相比较,视差图的精度提高了0.45%,且运行时间缩短了一半。

In order to solve the problem of the poor matching effect of stereo image in low texture and occlusion area in traditional stereo matching algorithm,this paper designed an end-to-end stereo matching algorithm based on convolutional neural network(CNN).The algorithm adopts residual convolutional neural network to extract the image features,and then uses the atrous spatial pyramid pooling(ASPP)module to obtain the context information of the image,and combines the multi-scale three-dimensional convolutional neural network to regularize the cost volume,and finally realizes the high-precision stereo matching algorithm.The mismatch rate of the obtained disparity map on the KITTI2015 testing platform was 2.42%.Compared with the Geometry and Context(GC)network,the accuracy of the disparity map was improved by 0.45%,as well as the running time was reduced by half.