Methods designed for traditional stereo matching problem are not suitable for micro-baseline stereo problem. In this paper, a novel phase-based learning framework is proposed dedicated to this problem. The inspiration comes from the relationship between phase in frequency domain and shifts in space domain. Steerable pyramid decomposition is used to compute the phase difference of micro-baseline stereo inputs, and learning based methods are adopted to determine disparity from phase difference patch. The proposed framework is a combination of domain transformation and machine learning, which exploits neighbor gradient information as well as data-driven benefits. Experimental results show that the combination effectively reduces the inherent error of phase-based methods, and our innovative framework outperforms traditional stereo matching methods.