Segment-Based Stereo Correspondence of Face Images Using Wavelets

In this paper, we introduce color segmentation based stereo correspondence for face images using wavelets. The intensity based correlation techniques are commonly employed to estimate the similarities between the stereo image pair, sensitive to shift variations and relatively lower performance in the featureless regions. Therefore, instead of pixel intensity, we consider wavelet coefficients of an approximation band, which is less sensitive to the shift variation. The approximation subband of reference image is segmented using mean shift segmentation method. A self-adapting dissimilarity measure that combines sum of absolute differences of wavelet coefficients and a gradient is employed to generate a disparity map of the stereo pairs. In our method instead of assigning a disparity value to a pixel, a disparity plane is assigned to each segment. Results show that the proposed technique produces smoother disparity maps with less computation cost.