Rectangular pixels for efficient color image sampling

Abstract

We present CFA designs that faithfully capture images with specified luminance and chrominance bandwidths. Previous academic research has mostly been concerned with maximizing PSNR of reconstructed images without specific regard to chrominance bandwidth and cross-talk. Commercial systems, on the other hand, pay close attention to both these parameters as well as to the visual quality of reconstructed images. They commonly sacrifice resolution by using a sufficiently aggressive OLPF to achieve low cross-talk and artifact free images.In this paper, we present the so called Chrominance Bandwidth Ratio, r, model in an attempt to capture both the chrominance bandwidth and the cross-talk between the various signals. Next, we examine the effect of tuning photosite aspect ratio, a hitherto neglected design parameter, and show the benefit of setting it at a different value than the pixel aspect ratio of the display. We derive panchromatic CFA patterns and corresponding photosite aspect ratios that provably minimize the photosite count for all values of r.An interesting outcome is a CFA design that captures full chrominance bandwidth, yet uses fewer photosites than the venerable color-stripe design. Another interesting outcome is a low cost practical CFA design that captures chrominance at half the resolution of luminance using only 4 unique filter colors, that lends itself to efficient linear demosaicking, and yet vastly outperforms the Bayer CFA with identical number of photosites demosaicked with state of the art compute-intensive nonlinear algorithms.

Disregarding spectral overlap — A unified approach for demosaicking, compressive sensing and color filter array design

“…The results are given in We also conducted robustness experiments by simulating sensor noise. The image reconstructed from the random RGB CFA suffered slightly lower increase in MSE due to noise as that from the Bayer CFA but the panchromatic design of [4] performed best suffering less than half as much increase in MSE.…”

“…We have experimentally verified rk(A) to be practically equal to |Ỹ |, for a random RGB CFA and over 300dB reconstruction using the said CFA with π/4 = 0.78 times as many photosites as that of the equivalent color stripe design. This is the theoretical lower bound on the CFA photosite count and is a significant improvement over the optimal no-spectral-overlap solution of [4] which needs 0.91 times as many photosites as the the color stripe CFA and has filters of many unique panchromatic colors.…”

“…This is in contrast to the non-overlapping spectra case wherein the CFA aspect ratio, and color pattern depend on the luminance to chrominance bandwidth ratio [4].…”

“…A quick succession of papers optimized CFA design techniques and presented improved CFA instances. The notable objectives include the minimization of pixel density by close packing of circular image spectra into the rectangular spectral support of the sensor [4,5,6] , optimization of the aspect ratio of sensor spectral support by controlling the aspect ratio of the photosites [4], minimization of overlap between modulated copies of image spectra [4,5,6] , color spaces that minimize chrominance energy [5,7] , quadrature modulation of image signals [4,5], maximization of CFA transmittance [5], maximization of demosaicker numerical stability [6], and avoidance of carriers with zero horizontal or vertical frequencies [3].…”