Adaptive Dual-Domain Debanding: A Novel Algorithm for Image and Video Enhancement
Kepeng Xu, Gang He, Tong Qiao, Zhenyang Liu · 2024
Banding artifacts substantially impair visual quality in images and videos post low-bitrate compression. Existing spatial domain approaches often inadequately address severe distortions due to limited receptive fields. In this work, we introduce an innovative adaptive debanding method that harmonizes operations across spatial and frequency domains, integrating a Spatial-Frequency Feature Extraction (SFFE) module and a Multi-Scale Spatial-Frequency Fusion (MSFF) module. Our algorithm adaptively mitigates high-frequency components associated with banding while preserving critical image details using targeted dynamic convolutions. Enhanced by multi-scale methods that amalgamate local and global priors, our approach efficaciously resolves color distortions. Comprehensive evaluations on the Band2k dataset demonstrate our method's superiority, achieving a PSNR of 40.04dB and an SSIM of 0.985, alongside a notable enhancement in the Deep Banding Index (DBI).