Digital camera spectral calibration

ISO 17321-2 describes procedures for characterizing digital cameras, which include determing scene analysis transforms using spectral measurements. This approach has a number of advantages over characterization methods which use a test chart comprising patches of known reflectance, from which a characterization model is usually calculated by regression. The spectral method computes a linear transform of normalised RGB values, using the camera spectral sensitivites and a training data set of measured spectral radiances.

It can be expected that the more representative the training data set is of the objects to be photographed, the smaller the characterization error will be. This is illustrated in the following set of results for two cameras:

Camera 1			Camera 2
ISO 17321	IE	Colorchecker	ISO 17321	IE	Colorchecker
1.44	1.92	3.16	1.96	2.11	2.79

In the table above, three different data sets were used to provide training data.

• ISO 17321-2 in-situ spectral radiance measurements for 25 natural objects
• Image Engineering test chart reflectances
• Macbeth ColorChecker spectral reflectances

The ISO 17321 data was used to test the characterization. The characterization errors shown are the average of the 25 colours in the ISO 17321 data set.

It can be seen that when the test set is the same as the training set (i.e. the ISO 17321 data is used for both) the errors are minimised, and increase when other data sets are used for training.

Other sources of spectral measurement data which can be used for general camera characterization include:

• ISO/TR 16066, Graphic technology - Standard object colour spectra database for colour reproduction evaluation (SOCS)

For most purposes a wide range of colours are needed to populate the training data. In applications where the range of object spectra encountered is more limited, it may be beneficial to use domain-specific training data. Examples of these include:

• Skin colour reflectance database [Data]
• etc

Some work has been done on selecting appropriate training sets for different applications, and on weighting the training set to optimise the result for a given application. ICC invites researchers who are interested in this topic, or who have results that can be shared, to contribute to the work. Please contact the ICC Technical Secretary.