Single Image Reflection Removal with Reflection Classifier and Gradient Restorer

Abstract

Abstract Techniques for removing image reflections can effectively eliminate artifacts superimposed on the subject due to factors such as reflections in light or through glass. Most current methods are designed on the assumption that reflection areas maintain the original image content. However, the neglect some extreme cases where there is no original information left, such as photos taken in museums with generated light spots.In this paper, we propose a novel model capable of further removing light spots reflections. Specifically, it takes an image with reflection contamination as input, and then guided by the proposed reflection classifiers and structure restorer, ultimately outputs a predicted transmission layer image.Experimental results demonstrate that the proposed model is applicable to different categories of reflection images, outperforming state-of-the-art reflection removal techniques.In summary, the proposed model improves the effect of image reflection technology based on artificial intelligence(AI) in the case of spot reflection.