2025 Volume 12 Issue 2 Pages 87-96
In this paper, we propose a novel bit depth enhancement (BDE) model that considers semantic contextual information by incorporating spatial feature transform (SFT) layers into the BDE model. In the proposed method, we adopt the pixel-wise class probability maps of the input image obtained by semantic segmentation as prior information for SFT. The SFT layer transforms from the input feature maps into the modulated feature maps by considering the contextual information modulated using affine parameters generated from the prior information. The proposed method considers the pixel-wise details by a proposed network that preserves spatial dimensions and the contextual information by incorporating the SFT layers conditioned with semantic information. Moreover, the proposed method adopts a perceptual loss function to recover the visually natural luminance changes by considering the contextual information. The experimental results show that the proposed BDE method achieves better performance compared with existing DNN-based BDE methods for restoring 8-bit depth from 3,4, and 6-bit depths. In addition, we investigate how to provide the contextual information to the BDE model, and show that providing it through the SFT layer is effective compared with other providing methods.
