Specular-based illumination estimation using blind signal separation techniques

Illumination estimation is important in many approaches to colour constancy, where object colour is measured without the effect of the spectral distribution of the illumination. Many illumination estimation methods for achieving colour constancy, particularly those based on the dichromatic reflection model, have performance limitations because they operate on images composed of blended specular and diffuse reflection components, and they may require image segmentation into regions; segmentation is a well-known image-processing challenge. This study proposes an illumination estimation method, which uses constrained blind signal separation anchored on the dichromatic reflection model, connected to a linear model of the illumination spectrum. Unlike conventional methods that use mixed-image components, the proposed method uses a specular image component extracted explicitly by blind signal separation. This can yield better illumination estimates, and blind signal separation can avoid image segmentation problems. Results of experiments show that the proposed method can recover the illumination spectral distribution, and that the extracted specular component yields better illumination estimation than mixed components. Similar results were observed for the two blind signal separation techniques assessed in this study; namely, the spatially constrained FastICA and independent component analysis based on mutual information.