Learning Colour Constancy Using Convolutional Neural Networks

Colour constancy has attracted attention of researchers from the academy and industry as it is a fundamental preprocessing task in many computer vision applications. Colour constancy is a feature of human visual system which enables humans to perceive colors of the objects invariant to the illuminant. However, it has been a challenging problem for computers due to its ill-posed structure. Arti cial neural networks have recently been very popular due to breakthrough results of deep neural networks in recognition tasks. Deep neural networks learn hierarchical representations (features) of data, which has started a new era in machine learning eld. Deep neural network models combine the feature learning and regression as a complete optimization procedure, namely they are an end-to-end learning approach. In this thesis, we investigate learning colour constancy using deep convolutional neural (CNN) networks. Unlike traditional color constancy methods, CNN model does not rely on any explicit imaging assumptions and hand-crafted features. Two di fferent CNN models are trained and evaluated on two widely used datasets (Shi-Gehler and SFU subset) from scratch. The results are compared with traditional statistics based approaches. It has been justi ed that CNN model signi cantly outperforms statistics based methods on both datasets. The improvements in average angular error are 26.6% and 20% for Shi-Gehler and SFU subset respectively