Removing reflection from images based on BP neural network of motion field

This paper introduces a method for removing reflection from images when imaging a scene behind glass surface. Rather than capturing a single image, we take advantage of an image sequence while moving the camera horizontal. By computing sparse motion fields between the reference image and all the other adjacent images, our method avoids to capture image sequence with rigid interval. We exploit the BP neural network algorithm to generate the dense motion field. It can quickly and accurately reconstruct the separated two images. Unlike previous approaches that exploit low-resolution image, we use the edge of images with high resolution to speed our compute. We demonstrate the method with some experiments. The result shows that our method works well on removing reflection. The method proposed makes the processing more efficient, and can be extended to the general.