Single Image Reflection Removal via Attention Model and SN-GAN

Single image reflection removal is of great practical importance for various computer vision tasks. Most non-learning methods try to solve this problem through the model-optimization scheme, which fails to produce promising results due to the shortage of suitable priors to model the difference between the reflection layer and the transmission layer. This paper presents an improved generative adversarial network to resolve this problem. First, we suggest that reflection removal is not only a channel-wise separation problem, but also a spatial variational occlusion removal task, which is sensitive to both spatial and channel-wise features. To this end, we integrate the CBAM module into the generator to enhance both spatial and channel-wise feature representation. Second, we consider the reflection layer as a spatial mask with space-relevant reflection intensity information, which can be used to elevate the performance of the discriminator. We then design a novel SNGAN structure with utilize the predicted reflection as a guidance to achieve better adversarial supervision. Specifically, our new generative network has an encoder-decoder structure with skip-connections, where the attention enhancement block is integrated into each skip-connection of the encoder-decoder subnet, and followed by an eight-layer fully convolutional subnet. Furthermore, the SNGAN loss is combined with L2 pixel loss and L1 VGG19 perceptual loss for training. The experimental results with benchmark datasets indicate that our method outperforms several state-of-the-art networks.