Significance map encoding and decoding using partition selection

What is this patent about?

’052 is related to the field of video data compression, specifically the entropy encoding and decoding of significance maps. In modern codecs like HEVC, significance maps identify the locations of non-zero coefficients within a transform unit after spectral transformation and quantization. Because these maps account for a large portion of the total bitstream, efficient context-adaptive binary arithmetic coding is essential for reducing data redundancy while maintaining high processing speeds.

The underlying idea behind ’052 is that traditional uniform partitioning of significance maps—where contexts are distributed evenly across a block—fails to account for the statistical concentration of energy in the low-frequency regions. The invention recognizes that bit positions in the upper-left area of a transform unit are used more frequently and carry different probability characteristics than those in the bottom-right. By employing non-spatially-uniform partitioning, the system can allocate more granular contexts to high-traffic areas and shared contexts to low-traffic areas, optimizing the balance between model accuracy and the speed of probability estimation convergence.

The claims of ’052 focus on a specific block-based mapping for 4×4 transform units that assigns nine distinct contexts across the fifteen relevant bit positions. This mapping is defined by a precise numerical sequence—0, 1, 2, 3, 4, 5, 2, 3, 6, 6, 7, 7, 8, 8, 7—which dictates how each coordinate in the block is associated with a context for entropy coding. This specific arrangement ensures that the most critical low-frequency coefficients receive dedicated or carefully grouped contexts, while higher-frequency positions share contexts to prevent the overhead of tracking underutilized probability models.

In practice, the invention operates by looking up the assigned context for each bit position during the scanning process. As the encoder or decoder traverses the block, it retrieves the current probability state for the assigned context, processes the bit, and immediately updates that state to reflect the new data. This context-adaptive approach allows the codec to learn the local statistics of the video slice dynamically. The implementation also supports switching between different partition sets based on the size of the encoded slice, ensuring that the complexity of the context model scales with the amount of data available to train it.

This approach differs from prior solutions by moving away from rigid, uniform grids that treat all coefficient positions with equal weight. By utilizing a refinement-based hierarchy, the invention allows a decoder to initialize a complex partition set using the states of a simpler, coarser set, facilitating a smooth transition as more data is processed. This mechanism effectively solves the problem of context dilution, where having too many contexts for too little data leads to poor probability estimation, thereby improving overall compression efficiency without significantly increasing the computational footprint of the entropy engine.