摘要
立体匹配是寻找立体图像对中对应点的问题,是立体视觉的核心问题。现有立体匹配算法通常是就立体匹配问题建立适当的数学模型并进行求解,在匹配速度和匹配精度之间存在矛盾。以生物视觉研究为背景,提出一种基于脉冲耦合神经网络(PCNN)的立体匹配方法。该方法以Markov随机场(MRF)上的贝叶斯模型为基础,并利用PCNN建立其似然概率模型。将左右2幅图像分别输入到2个PCNN网络,通过迭代生成点火时间序列。引入点火时间序列的平均点火时间差的概念,利用2个像素对应神经元的平均点火时间差来评价2个像素的相似性,并以此为基础确定似然概率。最后利用信任传递(BP)算法求解Markov随机场模型的最大后验概率问题。利用广泛使用的立体视觉测试图像对算法进行了实验。实验结果表明该算法能够有效实现立体匹配,匹配效果较好。
Stereo matching is the problem of finding corresponding pixels in two stereo images, which is the kernel problem of stereo vision. Existing stereo matching algorithms usually build a proper mathematical model for the stereo matching problem and get the solution. These kind algorithms have some disadvantages, and the conflict between matching accurate and matching speed exists. Taking biological vision study as a background, this paper proposes a stereo matching method based on pulse-coupled neural network (PCNN). This method is based on the Bayisian mod- el on the Markov random field (MRF) , and builds its likelihood probability model using PCNN. The left and right images are fed into two individual PCNN respectively, after iterations the PCNNs yield firing time series;and the concept of average firing time difference of the firing time series is introduced ; then the average firing time difference of the corresponding neurons of the two pixels is used to evaluate the similarity of the two pixels,based on which the likelihood probability is determined. Finally, the belief propagation (BP) algorithm is used to solve the maximum a posterior (MAP) problem of the MRF model. This method was tested using the widely used stereo vision test images,and the experiment results show that the proposed method can effectively achieve image stereo matching and obtain good matching results.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2013年第7期1539-1545,共7页
Chinese Journal of Scientific Instrument
基金
国家863计划(2011AA11A102)
"十二五"国家科技支撑计划(2011BAG01B04)
新世纪优秀人才支持计划(NCET-11-0572)资助项目
关键词
脉冲耦合神经网络
立体匹配
立体视觉
统计推断
随机场
pulse-coupled neural network
stereo matching
stereo vision
statistic inference
random field