Position Encoding Based Cross-Component Chroma Prediction for VVC

The correlation between components is a key characteristic of video data. To improve chroma prediction performance in VVC, attention-based cross-component prediction methods have shown promising results. However, most existing approaches adjust attention weights solely based on grayscale values, neglecting spatial position relationships and limiting spatial perception. To address this, we propose a position encoding-based cross-component chroma prediction algorithm. By leveraging the correlation variation of grayscale values with spatial distance, the algorithm integrates a learnable trigonometric position encoding scheme into the attention mechanism to explicitly model relative spatial relationships. An enhanced coordinate gating module is also introduced to strengthen cross-component feature representation. Experimental results show that, under constrained All Intra (AI) conditions, the proposed method achieves average BD-rate reductions of 0.59%, 3.61%, and 2.54% on the Y, U, and V components, respectively, compared to VTM.