A neural network approach to colour constancy

This thesis presents a neural network approach to colour constancy: a neural network is used to estimate the chromaticity of the illuminant in a scene based only on the image data collected by a digital camera. This is accomplished by training the neural network to learn the relationship between the pixels in a scene and the chromaticity of the scene's illumination. From a computational perspective, the goal of colour constancy is defined to be the transformation of a source image, taken under an unknown illuminant, to a target image, identical to one that would have been obtained by the same camera, for the same scene, under a standard illuminant. A colour constancy algorithm first estimates the colour of the illumination and second corrects the image based on this illuminant estimate. Estimating the illumination in a scene is a difficult task, since it is an inherently underdetermined problem.; Tests were performed on synthesised scenes as well as on natural images, taken with a digital camera. It is expected that theoretical models used for training that closely match the ‘real world’ lead to better estimates of the illuminant in real images. Thus, a natural step was to train the network on data derived from real images instead of synthetic scenes. This approach led to even more accurate estimates, of approximately 5ΔELab. To overcome the fact that the actual illuminant used in the training set images must be accurately known, and therefore must be measured for every image, a novel training algorithm called ‘neural network bootstrapping’ was developed. Experiments indicate that a grey world algorithm provides a relatively good estimation of the illuminant for images with lots of colours. This estimation, in turn, can be used for training the neural network. The final performance of the neural network is better than the performance of the grey world algorithm that was Initially used to train it.; The last part of the thesis deals with the issue of colour correcting images of unknown origin, such as images downloaded from the Internet or scanned from film. We have shown that colour correction of non-linear images can be done in the same way as for linear images and that a neural network is able to estimate the illuminant even when the sensor sensitivity functions and camera balance are unknown.; Using a neural network to estimate the chromaticity of the scene illumination improved upon existing colour constancy algorithms by an increase in both accuracy and stability. Therefore, neural networks provide a viable method for eliminating colour casts in digital photography and for creating illuminant-independent colour descriptors for colour-based object recognition systems.