Dense disparity estimation based on the bi-dimensional empirical mode decomposition and Riesz transformation

This paper proposed to apply the Bi-dimensional Empirical Mode Decomposition (BEMD) to the dense disparity estimation problem. The BEMD is a fully data-driven method and does not need predetermined filter and wavelet functions. It is locally adaptive and has obvious advantages in analyzing non-linear and non-stationary signals. Firstly we decompose the original stereo images by 2D-sifting process of the BEMD respectively. Through this procedure, a serial of Intrinsic Mode Functions (IMFs) and a residue are achieved. The residue denotes the DC component of the signal. Secondly, subtract the residue from original image. The resulting two dimensional signals can be thought of being free of disturbing frequencies, such as noise and illumination components. Subsequently, to obtain robust local structure information of the images, the plural Riesz transformation is utilized to achieve corresponding 2D analytic signals of the images. Thirdly, extract local phase information of the analytic signals. The similarity of local phase of stereo images, instead of local intensity information, are taken as the basis of calculating matching cost, which could reveal local structure with more robustness. At last, dense disparity map is estimated based on the proposed method. The winnertakes-all (WTA) strategy is applied to compute disparity of each pixel separately. Comparative experiment is conducted to compare the performance of the method with intensity-based methods. Rather good results have been achieved.