Kobus Barnard and Brian Funt, "Investigations into multi-scale retinex," in Colour Imaging: Vision and Technology, pp. 9-17, John Wiley and Sons (1999). [ Full text (pdf) ]
The main thrust of this paper is to investigate the processing effects of multi-scale retinex (MSR) and to provide a theoretical explanation of these effects. We first discuss possible objectives of such processing. We then study in detail the objective of improving images with respect to dynamic range and situations in which there is a mismatch between the scene illumination and appropriate illumination for the imaging system. This leads to a new algorithm that is more flexible, more accurate with respect to colour, computationally cheaper, and yet preserves the contrast-enhancement benefits of the original MSR method. To accomplish this we identify explicit and implicit benefits of MSR processing. We then decouple these operations from one another, and build an algorithm composed of independent steps that separates out the issues of gamma adjustment, colour balance, local dynamic range compression, global dynamic range compression, and colour enhancement, all of which can be interpreted as being part of the original MSR method. We posit that performing these operations independently makes it possible to do them more effectively. Furthermore, the overall algorithm is more flexible because each of the operations can now be independently tailored for a given application. We present an example of such an algorithm where colour constancy is provided by a neural network, and the colour is then preserved by a modification of MSR which performs dynamic range compression on the image luminance as opposed to the three channels separately. In this algorithm, the dynamic range compression is designed to match that of the original MSR method, but by using a single luminance channel as opposed to all three colour channels, the algorithm accomplishes the dynamic range compression at approximately one third the computational cost. Finally, we are careful to specify the algorithm with respect to linear input and output devices, thus achieving device independence.
Keywords: colour, color constancy, dynamic range compression, Retinex,
Full text (pdf)