Owing to the growing ubiquity of digital image acquisition and display, several factors must be considered when developing systems to meet future color image processing needs, including improved quality, increased throughput, and greater cost-effectiveness. In consumer still-camera and video applications, color images are typically obtained via a spatial subsampling procedure implemented as a color filter array (CFA), a physical construction whereby only a single component of the color space is measured at each pixel location. Substantial work in both industry and academia has been dedicated to post-processing this acquired raw image data as part of the so-called image processing pipeline, including in particular the canonical demosaicking task of reconstructing a full-color image from the spatially subsampled and incomplete data acquired using a CFA. However, as we detail in this chapter, the inherent shortcomings of contemporary CFA designs mean that subsequent processing steps often yield diminishing returns in terms of image quality. For example, though distortion may be masked to some extent by motion blur and compression, the loss of image quality resulting from all but the most computationally expensive state-of-the-art methods is unambiguously apparent to the practiced eye. … As the CFA represents one of the first steps in the image acquisition pipeline, it largely determines the maximal resolution and computational efficiencies achievable by subsequent processing schemes.

Here, we show that the attainable spatial resolution yielded by a particular choice of CFA is quantifiable and propose new CFA designs to maximize it. In contrast to the majority of the demosaicking literature, we explicitly consider the interplay between CFA design and properties of typical image data and its implications for spatial reconstruction quality.

Formally, we pose the CFA design problem as simultaneously maximizing the allowable spatio-spectral support of luminance and chrominance channels, subject to a partitioning requirement in the Fourier representation of the sensor data. This classical aliasing-free condition preserves the integrity of the color image data and thereby guarantees exact reconstruction when demosaicking is implemented as demodulation (demultiplexing in frequency).