Swin Transformer Based Network with Residual Channel Attention for Bit-Depth Expansion

抄録

Bit-depth expansion is a technique for reconstructing a high-bit image by predicting the missing bits in a low-bit image. With the development of high-bit monitors, corresponding high-bit images are needed to maximize their performance. However, many image data are still in 8-bit format. It is a complicated task to accurately recover lost information by expanding the bit depth while distinguishing between false contours and edges in real images. In this study, we propose a novel bit-depth expansion method using a Swin Transformer-based network with Channel Attention Layers (CALs). This network achieves high-performance bit-depth expansion by utilizing not only spatial features, which is an advantage of the Swin Transformer-based network, but also the correlation between channels obtained by CALs. Experimental results show that the proposed method outperforms conventional methods.