TSGAN: A Two-Stream Generative Adversarial Network for Bit-Depth Expansion

Abstract

The rapid development of display devices and the continuous improvement of visual quality requirements put urgent demands on high bit-depth (HBD) image, but most legacy data are stored and transmitted at low bit-depth (LBD). Bit-depth expansion (BDE) methods are designed to handle this dilemma by reconstructing HBD contents from LBD. But the expansion process often introduces artifacts such as false contours and color distortion. In this paper, we propose a two-stream generative adversarial network to expand bit-depth while suppressing artifacts. The generator network decomposes input image into base stream (BS) and detail stream (DS), then processes two data streams with different sub-networks respectively, and finally fuse two feature streams to reconstruct HBD image. Quantitative and qualitative experiments show that the proposed method outperforms tested BDE methods and CNN variants.