Context-guided Resolution Rescaling Framework for Intra Frame Coding

Deep learning techniques are increasingly integrated into rescaling-based video compression frameworks and have shown great potential in improving compression efficiency. However, existing methods achieve limited performance because 1) they often employ a uniform sampling ratio across regions with varying content complexity, resulting in the loss of important information; 2) these methods ignore the reuse of contextual information that captures texture and quality properties between regions, which may not effectively handle resampling and compression noises. To this end, this paper proposes a spatial context-guided resolution rescaling framework for intra frame coding, called CG-RRF, consisting of three sub-networks: a Texture-Quality context priors processor (CPP), a content-aware downscaling network (CDN), and a context-assisted upscaling network (CUN). Specifically, Texture-Quality CPP is designed to process context priors generated from codec, which are utilized to generate texture complexity map $T_{C}$, texture similarity score $T_{S}$ and quality-aware features for guiding the rescaling process. Then, CDN is proposed to utilize $T_{C}$ to separate the coarse-grained and fine-grained information for regions of varying complexity. Furthermore, CDN uses $T_{S}$ to adaptively adjust the proportion of fine-grained information, which can preserve essential information during downscaling. After that, CUN is designed to selectively aggregate high-quality context features guided by quality-aware features, and compensate current features with high-quality features based on $T_{S}$, which can suppress compression artifacts and generate a fine-grained high-resolution image. Extensive experiments show that our network achieves a significant 16.9% Bjøntegaard Delta Rate (BD-Rate) reduction under all-intra configuration compared to the codec anchor.