Reflection removal for stele images via sparse signal decomposition

Steles are most important historical records of human civilizations, and most of them are collected by museums. Stele images are usually captured with flash compensation to the low-light condition of museum, resulting in annoying reflection spots due to the non-lambertian reflectance of polished stone steles. Observing that the reflection spots are bright at the center but fade away from the center, we propose modeling reflection spots using the sum of Gaussian (SoG) functions. As the content of steles is often characters or drawings, the textures are mainly contours and edges, which suggest a sparse distribution on the gradient domain. Based on these observation, we propose a signal decomposition model for the removal of reflection spots, decomposing the stele image into three components: total variation (TV) component, SoG component, and reference component. The derived optimization is minimized via an alternating direction method (ADM). Experimental results show that our method achieves promising reflection removing results compared with other two segmentation algorithms.