Fast Single Image Reflection Removal Using Multi-Stage Scale Space Network

Abstract

Images captured in front of a glass obstruction often suffer from degradation due to the presence of reflections. These reflections can be classified as either high transmitted or low transmitted depending upon whether the captured image is dominated by either the background or the reflections respectively. Current approaches either aim to handle only high transmitted reflections or propose to train a unified neural network for addressing both kinds of reflections. However, using a single network to address different types of reflections is not very effective. Further, these methods are also computationally expensive and impractical to deploy on devices with limited resources such as smartphones. To address these challenges, we present a multi-stage pipeline for single image reflection removal within a scale space framework to address low and high transmitted reflections separately. Specifically, we treat the removal of low transmitted reflections that typically obscure the desired background as an inpainting challenge, while we handle high transmitted reflections using conventional techniques. We use specialized networks for these types of reflections within a scale space architecture that is light weight and is capable of removing reflections from very high resolution images. Our method shows superior performance both qualitatively and quantitatively compared to state of the art methods and our smartphone implementation takes about ~5 seconds to generate a high resolution 12 MP image.

Keywords

• Deep learning
• reflection removal
• computer vision