When imaging a scene through a glass,the image is a linear superposition of a real image observed through a glass and a virtual image reflected on it. This brings adverse effect on image processing,computer vision,etc. Reflective separation can be done by the original method of blind separation,nonlinearity in the course of imaging will degrade precision of separating reflections using linear independent component analysis. Kernel independent component analysis (KICA) can effectively deal with the nonlinearity. The reflections is removed by using FastKICA based on the Hilbert-Schmidt independence criterion. Experiments show that FastKICA is more effective than linear independent component analysis and conventional KICA.