High dynamic range (HDR) images are often claimed to contain accurate photometric and colorimetric measurements of the captured scenes. Since the HDR merging process removes the non-linearity introduced by a camera, the values are indeed approximately linearly related to luminance. However, there are also other factors that can distort captured color values, such as lens glare [MR07], vignetting, noise and spatially variant image processing of the camera. Moreover, the spectral sensitivity of most cameras sensors is different from the color matching functions (such as CIE XYZ 1931), therefore some inaccuracies in captured color can be expected. In this work we test how accurately HDR images can be color-calibrated. We capture HDR images using the multiexposure technique with camera RAW images. Because we use RAW images, there is mimimum camera image processing introduced in the captured images. The calibration is performed using a single HDR image containing an X-Rite color checker, in which color values have been measured using a photospectrometer. The accuracy of the calibration is tested in three other scenes, in which we placed diffuse color targets, which were also measured with the photospectrometer. The color calibration is tested for three cameras (Canon 1000D, Canon 550D and Panasonic Lumix DMC-LX7), and two camera calibration models.

Figure: Comparison of images before and after calibration for the “Library” scene

Figure: Comparison of images before and after calibration for the “Library” scene

Three cameras were used to take HDR photographs of 4 scenes containing a color checker chart (XRite) and custom color patches. These were also measured using a spectrometer to provide ground truth data. One of the scenes was then used as a training for 2 camera calibration models, one including and one without the black level. The parameters of the models were fitted by optimizing an error function based on the CIE2000 ∆E color error measure. For the remaining 3 scenes, the accuracy of the fit was verified using the CIE2000 ∆E error between the measured and predicted color values. Our results indicate that a very good fit can be achieved, in the range of 0.8 to 2.1 ∆E using the model without black level.

More information: Google Scholar