基于SIFT的双目图像深度信息提取
发布时间：2019-03-12 12:41
【摘要】：双目图像深度提取属于计算机视觉领域。本文研究目的是对一些已有的双目图像深度提取算法进行改进。在局部算法方面,利用SIFT特征组合绝对差值作匹配代价度量在保持效果的同时加快匹配速度;在全局算法方面,利用图像分割对置信传播作出效果方面的改进。具体的研究内容和改进方面如下:1.研究双目图像提取深度信息的原理,以及立体匹配的匹配代价度量方法和约束条件,约束条件使具有病态性的立体匹配问题变成理论上可解。在此基础上,研究基于窗口的匹配代价聚合方法,固定窗口法和自适应窗口法的多窗口法,并根据图像处理的模板加权重思想实现固定窗口法的一个改进,改进后效果与多窗口法接近,计算量近似是多窗口法的1/9。2.利用SIFT特征提取加快行传播法的计算速度。行传播法是基于颜色相似对图像作行线分割,再利用行线区域进行聚合的局部算法。原算法的匹代价度量是组合绝对差值和Census变换,因Census变换涉及大量逻辑比较运算而使计算时间长。针对这一点,本文利用SIFT特征提取代替Census变换来改进原算法,绝对差值和SIFT特征的组合度量在效果上与原度量接近,但计算时间减少近75%。3.利用图像分割改进置信传播的效果。置信传播是基于马尔可夫随机场来最小化全局能量函数的迭代算法。当迭代次数过多时,视差值平滑过度,造成图像中视差不连续区域效果变差。本文结合分水岭分割和Canny边缘提取阻止图像边缘传播避免平滑过度。迭代20次后,原算法、结合分水岭分割和结合Canny边缘提取的算法的误匹配率分别是4.79%、3.86%、3.25%,效果得到提升。
[Abstract]:Depth extraction of binocular images belongs to the field of computer vision. The purpose of this paper is to improve some existing depth extraction algorithms for binocular images. In the aspect of local algorithm, the absolute difference of SIFT feature combination is used to measure the matching cost while maintaining the effect while accelerating the matching speed. In the aspect of global algorithm, the image segmentation is used to improve the effect of confidence propagation. Specific research contents and improvements are as follows: 1. In this paper, the principle of extracting depth information from binocular images is studied, and the matching cost measurement method and constraint conditions for stereo matching are studied. The constraint conditions make the ill-conditioned stereo matching problem theoretically solvable. On this basis, the window-based matching cost aggregation method, fixed window method and adaptive window method of multi-window method are studied, and an improvement of fixed window method is realized according to the idea of template weighting in image processing. The result of the improved method is close to that of the multi-window method, and the calculation amount is approximately 1 / 9.2 of the multi-window method. SIFT feature extraction is used to speed up the calculation of line propagation method. Line propagation method is a local algorithm based on color similarity for line segmentation and aggregation of line regions. The computation cost measure of the original algorithm is the combination absolute difference and Census transform. Because the Census transform involves a large number of logical comparison operations, the computation time is long. In view of this, this paper uses SIFT feature extraction instead of Census transform to improve the original algorithm. The combined measurement of absolute difference and SIFT feature is close to the original measure in effect, but the calculation time is reduced by nearly 75%. Image segmentation is used to improve the effect of confidence propagation. Confidence propagation is an iterative algorithm based on Markov random fields to minimize the global energy function. When the number of iterations is too many, the parallax smoothing is excessive, which causes the disparity discontinuity in the image to become worse. In this paper, we combine watershed segmentation and Canny edge extraction to prevent image edge propagation from smoothing. After 20 iterations, the error matching rate of the original algorithm, which combines watershed segmentation and Canny edge extraction, is 4.79%, 3.86% and 3.25%, respectively.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
本文编号：2438784
[Abstract]:Depth extraction of binocular images belongs to the field of computer vision. The purpose of this paper is to improve some existing depth extraction algorithms for binocular images. In the aspect of local algorithm, the absolute difference of SIFT feature combination is used to measure the matching cost while maintaining the effect while accelerating the matching speed. In the aspect of global algorithm, the image segmentation is used to improve the effect of confidence propagation. Specific research contents and improvements are as follows: 1. In this paper, the principle of extracting depth information from binocular images is studied, and the matching cost measurement method and constraint conditions for stereo matching are studied. The constraint conditions make the ill-conditioned stereo matching problem theoretically solvable. On this basis, the window-based matching cost aggregation method, fixed window method and adaptive window method of multi-window method are studied, and an improvement of fixed window method is realized according to the idea of template weighting in image processing. The result of the improved method is close to that of the multi-window method, and the calculation amount is approximately 1 / 9.2 of the multi-window method. SIFT feature extraction is used to speed up the calculation of line propagation method. Line propagation method is a local algorithm based on color similarity for line segmentation and aggregation of line regions. The computation cost measure of the original algorithm is the combination absolute difference and Census transform. Because the Census transform involves a large number of logical comparison operations, the computation time is long. In view of this, this paper uses SIFT feature extraction instead of Census transform to improve the original algorithm. The combined measurement of absolute difference and SIFT feature is close to the original measure in effect, but the calculation time is reduced by nearly 75%. Image segmentation is used to improve the effect of confidence propagation. Confidence propagation is an iterative algorithm based on Markov random fields to minimize the global energy function. When the number of iterations is too many, the parallax smoothing is excessive, which causes the disparity discontinuity in the image to become worse. In this paper, we combine watershed segmentation and Canny edge extraction to prevent image edge propagation from smoothing. After 20 iterations, the error matching rate of the original algorithm, which combines watershed segmentation and Canny edge extraction, is 4.79%, 3.86% and 3.25%, respectively.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
【参考文献】
相关期刊论文 前3条
1 白明;庄严;王伟;;双目立体匹配算法的研究与进展[J];控制与决策;2008年07期
2 慕春棣,tsinghua.edu.cn,戴剑彬,叶俊;用于数据挖掘的贝叶斯网络[J];软件学报;2000年05期
3 匡锦瑜;吉布斯随机场模型及其在图象处理中的应用[J];通信学报;1990年03期
相关博士学位论文 前2条
1 施陈博;快速图像配准和高精度立体匹配算法研究[D];清华大学;2011年
2 郑志刚;高精度摄像机标定和鲁棒立体匹配算法研究[D];中国科学技术大学;2008年
，本文编号：2438784
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2438784.html
