Camera Color Correction Using Two-Dimensional Transforms

For digital camera systems, transforming from the native camera RGB signals into an intermediate working space is often required, with common examples involving transformations into XYZ or the sRGB. For scene-linear camera signals, by far the most common approach utilizes 3x3 matrices. For color pipelines designed for Rec709 displays, matrix-based input transforms are capable of producing reasonable accuracy in this domain. However, the associated colorimetric errors can become significant for saturated colors, for example those beyond Rec709. To address this shortfall, a novel input color transformation method has been developed that involves the use of twodimensional lookup tables (LUTs). Because the surfaces associated with the 2D LUTs possess many degrees of freedom, highly accurate colorimetric transformations can be achieved. For several cinematic and broadcast cameras tested, this new transformation method consistently shows a modest reduction of mean deltaE errors for lower-saturation colors. The improvement in accuracy becomes much more significant as saturation increases, such that the mean deltaE errors are reduced by more than a factor of three. Introduction Recently, several factors have contributed to wider adoption of digital capture systems in the professional environment, for example in cinematic and episodic productions. Among these include improved noise performance, extended dynamic range capability, and importantly the creation of cost-effective digital workflow ecosystems. As productions continue to migrate toward digital, lower cost camera systems have been introduced, giving many small to medium budget productions access to high quality content creation. For example, both the Canon C500 and the RED Epic are relatively affordable in the $20K-$30K range; even prosumer cameras, such as the GoPro Hero3 or the Canon 5D mII/mIII, have been successfully used in numerous productions [1,2]. Although at higher price points, the ARRI Alexa and Sony F55/F65 have both found widespread use, and have produced imagery with quality that rivals that of modern cinematic film. The image formation process for these systems is illustrated in Figure 1, for single sensor designs and more complex cameras utilizing multiple sensors. In the case of a single sensor configuration, a scene is imaged through the optical system onto the sensor. A color filter array (CFA) is patterned onto the sensor, and in the case of a Bayer design produces individual pixels with either a red, green, or blue response. With this CFA design, the spatial sampling of the green pixels is twice that of the red or blue channels, and to produce separate red, green, and blue images with the full sensor pixel count, various demosaicing algorithms are employed [3]. For three chip configurations (typically found in broadcast camera systems), dichroic mirrors in conjunction with red, green, and blue trimming filters produce the full resolution RGB channels without the need for demosaicing [4]. The spectral transmittance of the color filters when combined with the absorption of silicon form the spectral response characteristics of each channel. Example curves for single chip and multi-sensor cameras are shown in Figure 2. Analog RGB signals, initially in the form of electrons in the well of the photodiode associated with each pixel, are formed by taking the projection of the focal plane spectral power distribution L() and the RGB spectral sensitivity functions ̅ , ̅ , over all wavelengths: