Video Transform Coefficient Sign Prediction Along Skip Directions

The patent applies different entropy coding strategies based on the specific intra prediction direction. Statistically more likely directions (e.g., horizontal and vertical) use optimized context models with fewer bits, while less likely directions use different context assignments. This local differentiation reduces overall bit usage without requiring a completely complex system, as each direction gets tailored treatment based on its statistical probability.

The patent changes the context parameters (context models) used for entropy coding based on the intra prediction direction. By selecting different context assignments for different directions, the system adapts the coding parameters to match the statistical characteristics of each direction, thereby reducing bit usage for less likely directions while maintaining decoding simplicity through predefined context tables.

The patent implements dynamic context selection where the entropy coding context changes based on the intra prediction direction. Instead of using a static coding scheme for all directions, the system dynamically selects appropriate context models from predefined sets, allowing more efficient bit representation for each direction while maintaining the ability to accurately represent all possible directions through the diverse context options.

The patent pre-defines multiple context models and their assignments for different intra prediction directions before encoding begins. This preliminary preparation allows the encoder to quickly select the appropriate context for each direction without complex real-time calculations, thereby maintaining coding accuracy while improving compression efficiency through optimized bit usage for each direction.

The patent employs sign value prediction where the sign of a transform coefficient is predicted from previously decoded neighboring coefficients. This self-service approach reduces redundancy by exploiting the statistical correlation between neighboring coefficients, allowing the current coefficient's sign to be inferred from context rather than explicitly coded, thereby reducing bit usage without significantly increasing decoding complexity through simple prediction logic.

The patent uses feedback from previously decoded transform coefficients to determine the coding context for current coefficients. The decoded sign values of neighboring coefficients feed into the context selection process for subsequent coefficients, creating a feedback loop that reduces redundancy by adapting the coding strategy based on actual decoded values rather than assuming uniform distribution across all coefficients.