基于小波的多分辨率相位匹配方法研究

In computer vision, correspondence point matching is one of the most important key research branches. According to the difference among matching elements, there are three different approaches as area based matching, feature based matching and phase based matching. Above these three different matching approaches, phase matching is done in frequency domain and developed in recently twenty years.In this project, it is studied for the phase based correspondence point matching. We mainly focused on the generalized approach of phase matching, and also the features in the approach of phase matching. According to the features of phase matching, the disparity is computed from coarse-to-fine. This makes the approach of phase matching to be multiresolution. We introduced the wavelet, which has multiresolution decomposition properties, into the phase matching approach. In our wavelet based phase matching approach, the local phase is computed by applying Hilbert transform to multiresolution wavelet decomposed images. Such approach not only solves the contradiction between the coarse-to-fine disparity estimation approach and non-multiresolution local phase computation existed in Gabor based approaches, but also improves the disparity computation efficiency.In phase matching, there exists singularity in local phase. This makes the disparity estimated not so robust at where singularity existed. In this project, the robust disparity estimation algorithms are also studied. According to Phase Congruency theory, there exists a close relationship between image features and local energy. Based on this theory, we proposed an improved phase matching algorithm with more robust to phase singularity. And also in phase matching algorithm, we adopted dynamic programming into phase matching, and designed a cost function for disparity estimation in phase domain. The proposed algorithm combined local phase with image's intensity to estimate disparity and detect occlusion at the same time.

本项目将小波变换及多分辨率分析方法引入到基于相位匹配的双目计算机视觉中，研究基于小波变换的局部相位的多分辨率提取及匹配方法，讨论小波变换条件下多分辨率相位匹配快速实现算法。研究的算法在实现每点亚象素级匹配精度条件下，具有很高的匹配速度、很强的鲁棒性。有关问题的研究将为图象匹配技术引入新的思想方法和工具，算法实用可靠。