Disparity Estimation From Stereo Images With Multilayered Reaction-Diffusion Models Of Activation-Inhibition Mechanism

The present paper proposes a method that estimates a stereo disparity map. The method consists of multilayered networks of reaction-diffusion equations having activator and inhibitor variables. A particular set of equations describes non-linear oscillators coupled with diffusion processes, and governs a disparity layer. Estimating a stereo disparity map requires the two main constraints: the smoothness constraint and the uniqueness. The activation process of the diffusion-coupled oscillators performs the filling-in process for a stereo disparity map as the smoothness constraint. The mutual inhibition mechanism among the multilayered networks retains the uniqueness constraint of the disparity at a particular pixel site. By applying the proposed equations to cross-correlation functions derived from stereo images, we obtain a self-organized stereo disparity map. Experimental results show that the proposed method is superior to a previous one, in particular, for real stereo images