Single Image Reflection Removal Based on Deep Residual Learning

Zhixin Xu, Xiaobao Guo, Guangming Lu

Abstract: We present a deep residual learning approach to address the single image reflection removal problem. Specifically, residual learning exploits the mapping between the observed image and its comparatively simple reflection information, which is then removed to obtain a clear background. Different from other methods that roughly eliminating the reflections and producing the images with remanent sticking, a novel generative adversarial framework is proposed, where the generator is embedded with the deep residual learning, significantly boosting the performance without impairing the intactness of the background by adversarial training. Moreover, a multi-part balanced loss is introduced with comprehensive consideration on the measure of feature similarity as well as the discriminating ability of GAN. It produces the result of high quality by learning the reflection and the background feature simultaneously. Experiments show that the proposed method achieves a state-of-the-art performance.