Learning filtered color planes for depth-extraction in coded aperture images

Coded aperture imaging presents an interesting approach to single-image depth extraction. Existing approaches use depth-dependent color misalignment brought about by color coded apertures to calculate the depth map of an image. The depth map itself is not directly calculated from the image but is taken by estimating the disparity for a given shift in the color channels. Earlier measures of color disparity utilized eigenvalue calculations over a sliding window. Such a technique, while effective, is also computationally expensive. In a previous work, we demonstrated that the disparity may be approximated using a variance calculation over a sliding window taken in the YCbCr color space. In this work, we further improve the approximation by learning a suitable filtered color plane that quickly captures the relevant information for disparity estimation. Simulations show the substantial improvement of the proposed color space over eigenvalue-based techniques and the YCbCr space while maintaining low computational cost.