Transmission Removal from a Single OSN-Shared Glass Mixture Image

Abstract

Photographs taken through glass often reflect the photographer or the surroundings, which is very helpful in uncovering information about the photograph. Various lossy operations performed on images over online social networks (OSNs), such as compression and resampling, pose a great challenge for transmission layer removal. This paper proposes a self-attention-based architecture for image enhancement over OSNs, to ensure that the downloaded glass mixture image can show more information about the reflection layer than the original image. Transmission layer removal is then achieved using a two-stage generative adversarial network. We also add attention to the transmission layer in the mixture image and use the gradient and color block information in the next stage to extract the reflection layer. This method yielded a gain of 0.46 dB in PSNR, 0.016 in SSIM, and 0.057 in LPIPS, resulting in an effective improvement in the visual quality of the final extracted reflection layer images.