Polarization reflection-transmission separation based on multi-feature fusion and adaptive segmentation

Imaging through glass is plagued by obstructive reflections. This paper introduces a novel polarization-based method to separate reflections from the desired background transmission. Our approach overcomes key limitations of existing techniques when dealing with non-planar glass and complex lighting. We achieve this by fusing multiple image features to create a complexity map, which guides an adaptive segmentation of the image into localized patches. This allows for independent optimization within each patch using a mutual information minimization criterion, effectively handling spatial variations across curved surfaces. A final weighted fusion ensures seamless results. Experiments demonstrate superior performance in recovering clear background imagery through challenging surfaces like windshields, outperforming current methods in both visual and quantitative measures.