利用视觉相位鉴别能力求解立体视觉匹配

Utilizing the Ability of Human Visual Phase Discrimination for Stereo Vision

本文采用图象相位信息作为立体匹配基元，它同时具有区域相关法的定位精度高，恢复视差密度大和特征匹配法的实现速度快等优点。为了增强算法处理复杂场景的能力，在算法设计时充分考虑了视差的不连续信息和遮掩问题的处理。实验结果表明，该方法具有较高的测量精度，基本上不需要内插等后续处理即可得到景物表面较完整的描述。

The motivation of this study is to explore the ability of human visual phase discrimination for stereo vision. A very important issue is to understand how much information is carried by the phase information in human stereo interpretation, hence helping for creating new models and approaches which are efficient for interpreting natural scence. Based on the visual cue, we have designed a new approach of stereo vision utilized the wavelet theory. Unlike the area-based that is computationally expensive, and also the feature-based which generates only sparse disparity map, the approach uses the image phase as the matching primitives which can generate a dense disparity field efficiently. In order to deal with complicated natural scenes, the constraints controlling the surface discontinuity and occlusion are also introduced in our approach. Experiments have proved that the proposed approach can be adaptable for almost arbitrary images and resulting disparity surface is accurate and dense. From this study we could probably understand what role of phase information is in human stereo fusion and help developing appropriate computational approach for practices.