Significance Map Encoding And Decoding Using Partition Selection

What is this patent about?

’783 is related to the field of video compression, specifically the encoding and decoding of significance maps within video codecs like H.264/AVC and HEVC. Significance maps indicate the locations of non-zero coefficients within a transformed block of video data, and their efficient encoding is crucial for overall compression performance. The patent addresses the problem of optimizing context modeling for entropy encoding these maps.

The underlying idea behind ’783 is to improve the efficiency of context-adaptive entropy coding of significance maps by using non-uniform partitioning of the map. Instead of assigning a unique context to each coefficient position or using uniform partitioning, the invention groups coefficient positions into parts based on their statistical significance, assigning a single context to each part. This allows for better context adaptation and improved compression, especially by concentrating contexts in the upper-left quadrant where significant coefficients are more likely to occur.

The claims of ’783 focus on a method, encoder, and non-transitory medium for encoding a significance map. The key element is a partition set that assigns contexts to bit positions in the significance map. This assignment is characterized by: (1) each bit position in the upper left quadrant having one of at least four contexts, (2) the upper right and lower left quadrants sharing four to six contexts different from the upper left, with each context assigned to at least two bit positions, and (3) the lower right quadrant having a majority of bit positions assigned to a single context.

In practice, the encoder iterates through each bit position in the significance map, determines the appropriate context based on the pre-defined partition set, encodes the bit value using that context, and then updates the context based on the encoded bit. The non-uniform partitioning strategy acknowledges that high-frequency coefficients (typically in the lower right) are often zero, so fewer contexts are needed there. This allows the encoder to concentrate its context modeling resources on the more statistically significant upper-left quadrant, leading to better probability estimation and improved compression.

This approach differentiates itself from prior methods that use either a uniform distribution of contexts or a one-to-one mapping of contexts to coefficient positions. By strategically grouping coefficient positions into parts and assigning contexts based on their statistical likelihood of containing significant coefficients, ’783 achieves a better balance between adaptivity and accuracy in context modeling. This results in a more efficient encoding of the significance map and, consequently, improved overall video compression performance.