AN EFFICIENT SINGLE IMAGE REFLECTION REMOVAL ALGORITHM WITH VISUAL ATTENTION MECHANISM

10.6329/CIEE.202312_30(2).0004

Yu-Jen Wei；Zhao-Rong Chen；Tien-Ying Kuo

deep learning ； reflection removal ； generative adversarial network (GAN) ； attention

電機工程學刊

30卷2期（2023 / 12 / 01）

71 - 80

英文

When a scene to be photographed is behind the glass medium, the photo often suffers from reflection artifacts, which compromise the integrity of the image content. Existing methods for solving this problem are ineffective and frequently necessitate the processing of multiple images taken from different camera angles as input. Therefore, we propose a learning-based algorithm for removing reflections that requires only a single image with reflections as input to produce a reflection-free image. Our method utilizes a reflection attention module that generates a probability map of reflection from latent features. A convolutional LSTM unit is used in this module to refine the generated probability maps. The output is then passed to a reflection removal module to produce a reflection-free image. We adopt the Patch-GAN framework to train the model so that the generated results are closer to real images. Our results outperform existing algorithms in several evaluation methods and provide a better visual experience.