Camera Sensor Invariant Auto White Balance Algorithm Weighting

Color constancy is the ability to perceive colors of objects, invariant to the color of the light source. The aim for color constancy algorithms is first to estimate the illuminant of the light source, and then correct the image so that the corrected image appears to be taken under a canonical light source. The task of the automatic white balance (AWB) is to do the same in digital cameras so that the images taken by a digital camera look as natural as possible. The main challenge rises due to the illposed nature of the problem, that is both the spectral distribution of the illuminant and the scene reflectance are unknown. Most common methods used for addressing the AWB problem are based on low-level statistics assuming that illuminant information can be extracted from the image’s spatial information. Nevertheless, in recent studies the problem has been approached with machine learning techniques quite often and they have been proved to be very useful. In this thesis, we investigate learning color constancy using artificial neural networks (ANNs). Two different artificial neural network approaches are utilized to generate a new AWB algorithm by weighting some of the existing AWB algorithms. The first approach proves to be better than the existing approaches in terms of median error. On the other hand, the second method, which is better also from system design point of view, is superior to others including the first approach in terms of mean and median error. Furthermore, we also analyze camera sensor invariance by quantifying how much the performance of the ANNs degrade when the test sensor is different than the training sensor