Joint Demosaicking / Rectification of Fisheye Camera Images using Multi-color Graph Laplacian Regulation

To compose one 360 degrees image from multiple viewpoint images taken from different fisheye cameras on a rig for viewing on a head-mounted display (HMD), a conventional processing pipeline first performs demosaicking on each fisheye camera's Bayer-patterned grid, then translates demosaicked pixels from the camera grid to a rectified image grid. By performing two image interpolation steps in sequence, interpolation errors can accumulate, and acquisition noise in each captured pixel can pollute its neighbors, resulting in correlated noise. In this paper, a joint processing framework is proposed that performs demosaicking and grid-to-grid mapping simultaneously, thus limiting noise pollution to one interpolation. Specifically, a reverse mapping function is first obtained from a regular on-grid location in the rectified image to an irregular off-grid location in the camera's Bayer-patterned image. For each pair of adjacent pixels in the rectified grid, its gradient is estimated using the pair's neighboring pixel gradients in three colors in the Bayer-patterned grid. A similarity graph is constructed based on the estimated gradients, and pixels are interpolated in the rectified grid directly via graph Laplacian regularization (GLR). To establish ground truth for objective testing, a large dataset containing pairs of simulated images both in the fisheye camera grid and the rectified image grid is built. Experiments show that the proposed joint demosaicking / rectification method outperforms competing schemes that execute demosaicking and rectification in sequence in both objective and subjective measures.