Semantic Guided Single Image Reflection Removal

Reflection is common when we see through a glass window, which not only visual disturbance but also influences the performance of computer vision algorithms. Removing reflection from single image, however, highly ill-posed since color at each pixel needs to be separated into two values belonging clear background and reflection, respectively. To solve this, existing methods use additional priors such as layer smoothness, double effect, consistency distinguish layers. However, these low-level may consistently valid in real cases. In this paper, inspired by fact that human beings can separate layers easily recognizing objects understanding scene, propose object semantic cue, high-level information, guidance help removal. Based on data analysis, develop multi-task end-to-end deep learning method with component, removal segmentation jointly. Extensive experiments different datasets show significant gain using object-oriented information. We demonstrate application our other tasks.