Removing achromatic reflections from color images with application to artwork imaging

We propose a general approach to remove reflections from a color image acquired through a semi-transparent medium, and show its application to the restoration of images of paintings framed behind glass and manuscripts laminated for conservation purposes, affected by the reflection of a light source. The problem is modeled by assuming that the unwanted reflection is an achromatic or monochromatic image that combines additively with the real transmitted image of the object of interest. In the absence of information about the mixing coefficients, we adopted a blind source separation technique that exploits the dependence of the three color channels of the original image, and the independence of the reflected image. In particular, these constraints are forced on the image gradients rather than on the intensity images. The algorithm is constituted of a step for the estimation, via independent component analysis of the model parameters, followed by a regularization technique to estimate the component images. The algorithm is very fast and provides promising results.