Reflection Separation Using Patch-Wise Sparse and Low-Rank Decomposition

This paper introduces a robust method for removing objectionable reflection interference in photographs captured through a piece of transparent medium. We exploit the fact that a group of image patches extracted from multiple correlated images with similar transmission lie in a very low-dimensional subspace, leading to a low-rank matrix after patch assembly. This allows us to formulate reflection separation as a per-patch sparse and low-rank decomposition problem which can be well solved by the ALM-ADM strategy. To eliminate the influence of unwanted reflection in patch searching and ensure that the extracted patches has a high similarity regarding their transmission layers, we introduce a new patch similarity metric based on both image intensities and gradients. This improves the performance of reflection separation. In addition, since our method does not require image reconstruction from gradient, color-shifting artifacts can be significantly ameliorated and more scene details can be preserved. Experimental results on both synthetic images and various real-world examples demonstrate that the proposed method achieves high quality reflection separation and performs favorably against many existing techniques.