Visualizing removed grime and varnish layers in painting restoration

We present a novel method for estimating and visualizing the grime and varnish layers in paintings using only images taken before and after their physical restoration. These layers offer valuable insight into the painting’s conservation history – grime indicates surface dust exposure, while varnish suggests aging or sunlight-induced deterioration. For each pixel in the before and after images, we estimated the reflectance spectrum using a neural network trained to convert RGB values into spectral data across the visible range. To isolate the dirt layer – conceptualized as a transmitting removed during restoration – we calculated point-per-point transmittance spectrum as the ratio of the pixel-wise reflectance between the dirty and cleaned versions. Since this layer comprises both surface grime and aged varnish, we further decomposed it into these components using spectral unmixing. Our mixing model follows the Beer-Lambert law which describes light transmittance through a medium as generally exponential. Accordingly, we expressed the transmittance of the dirt layer as the product of the transmittance of grime and varnish, each raised to an adjustable exponential weight representing their contribution. Our approach generated realistic and interpretable grime and varnish layers. These visualizations document what was removed during restoration and reveal patterns of dust buildup and UV exposure. Such insights help conservators assess past surface conditions and support informed decisions on preventive care, display, and long-term preservation, marking a new direction in digital art conservation.