CU-Level Dynamic Approximate Storage Scheme for VVC Intra-Frame Prediction

Abstract

This paper presents a CU-level dynamic approximate storage scheme for intra-frame prediction in Versatile Video Coding (VVC) encoders, targeting energy-efficient video compression. The proposed approach leverages the spatial-temporal correlation of intra-coded regions across neighboring Random Access (RA) frame layers to guide fine-grained approximation decisions. By exploiting intra-prevalence maps extracted from previously encoded frames, the scheme dynamically adjusts the approximation strength at the coding unit (CU) level. Experimental results demonstrate that the proposed method effectively mitigates coding-efficiency losses while preserving most of the energy savings afforded by approximate storage. On average, the CU-level refinement reduces BD-Rate penalties compared to baseline strategies (0.12-2.05%), especially for higher-resolution sequences (up to 5.02%), while introducing only a small energy overhead, thereby achieving structural enhancements in this trade-off optimization. These results indicate that fine-grained approximation guided by intra correlation is a promising direction for improving the trade-off between energy consumption and coding efficiency in VVC encoders.