The Step Where a Video Actually Loses Detail A Codec Claim Written All the Way to Quantisation

This application claims a video decoding method that runs the full lossy chain from end to end. A square block is quad split into four equal squares, one of those is sliced by an index into four thin strips or two halves, each piece is predicted with the same intra prediction mode, and then claim 1 does something its three same-day siblings do not: it recites obtaining a quantized transform coefficient, inverse-quantising it, and only then inverse transforming the result into the residual. It was filed by B1 Institute of Image Technology, Inc. of South Korea, with Ki Baek Kim as sole inventor.

The Problem

Prediction does not compress anything by itself. Nor does a transform. Both are rearrangements: a prediction replaces a block of pixels with a block of errors, and a transform replaces that block of errors with a block of frequency coefficients. Run those two steps and you still have as many numbers as you started with, and every one of them is exact. The file has not got smaller.

Compression happens when numbers get thrown away, and in a video codec that is the job of one step. Quantisation divides each transform coefficient by a step size and keeps only the whole-number part. Coefficients that were small become zero, coefficients that were large lose their least significant detail, and the resulting run of zeros costs almost nothing to encode. It is also irreversible. Everything a viewer notices as compression artefacts, from blockiness to smeared texture, traces back to the arithmetic in that one operation. Every other stage in the pipeline is either exactly invertible or a shared guess that the decoder can repeat.

The specification is unusually direct about how conditional the step is. Quantisation may be skipped according to the encoding configuration, it says, giving a quantisation parameter of zero, that is, a lossless compression environment, as one example, and the case where the characteristics of an image mean quantisation buys no compression as another. It can be skipped over part of a block rather than all of it, and the skip can be set explicitly or left implicit. The specification also notes that a residual may be transformed without being quantised, quantised without being transformed, or both, depending on configuration. In other words, a claim that names quantisation is naming a step the same document treats as optional.

Sitting underneath all of this is the reason the residual is small enough to be worth the trouble. The specification’s stated objects are intra prediction, subblock-wise intra prediction, and the partitioning and coding order of those subblocks. Splitting a block into strips and coding them one after another shortens the distance from each strip to the reconstructed pixels it predicts from, and sharing a single prediction mode across the strips stops the split from multiplying the signalling overhead. A better prediction means a weaker residual, which means the quantiser has less to destroy for a given bit budget.

What This Invention Does

Claim 1 is a decoding method with seven steps rather than the five its siblings use. It begins by generating a coding block using tree-based partitioning, defined in the claim as splitting a square block into four square blocks of the same size. It then generates sub-blocks by index-based partitioning, and the claim spells out two alternatives. A first partitioning splits a 4Nx4M coding block into four identical 4NxM sub-blocks or four identical Nx4M sub-blocks, N being equal to or different from M. A second partitioning splits the coding block into two equal parts, horizontally or vertically. A closing clause makes the choice automatic: whether to perform the first or the second partitioning is determined based on the size of the coding block.

Index-based partitioning is distinct from the tree splits in the specification’s own terms. A tree split derives its type from several pieces of information, while an index split picks a type straight out of a candidate group using one index value, and the specification restricts it to a single layer, meaning a sub-block produced this way is not split again.

Intra prediction then runs on each sub-block, subject to the requirement that the same intra prediction mode is used for all of them. At this point the claim diverges from its siblings. It obtains a quantized transform coefficient for the coding block. It obtains a transform coefficient by performing inverse-quantization on that quantized coefficient. It generates the residual block, and a wherein clause specifies that the residual block is generated by inverse-transforming the transform coefficient, not merely by inverse transforming information about the residual. Finally it reconstructs the coding block from the prediction block and the residual block.

Claim 5 mirrors this from the encoder side in the order the encoder actually works: generate the residual from the prediction block, generate a transform coefficient by transforming the residual, generate a quantized transform coefficient by performing quantisation on that coefficient, and encode the coding block by encoding the quantized transform coefficient. Claim 7 adds the transmission step, encoding the quantized transform coefficient into a bitstream and transmitting it. The dependent claims are modest by comparison: claim 2 conditions whether index-based partitioning is used at all on a property of the coding block, claim 3 says that property may be area, width or height, and claim 4 allows the partitioning type to be determined by a signalled index or by a property of the block.

Key Features

• Quantisation named inside the independent claim. Claim 1 recites obtaining a quantized transform coefficient and performing inverse-quantization on it, which is the limitation that separates this application from the three filed alongside it.
• A residual built from a named coefficient. The wherein clause requires inverse transforming the transform coefficient produced by the inverse-quantisation step, tying the two operations together rather than treating the residual as an unspecified input.
• Two index-based partitionings selected by block size. A 4Nx4M block can be cut into four identical strips in one direction, or the block can be halved, with the coding block’s size deciding which applies.
• One shared intra prediction mode. All sub-blocks are predicted with the same mode, which keeps the cost of splitting low while giving each strip closer reference pixels than the parent block would have had.
• A matching encoder chain. Claim 5 puts the steps in encoding order, transform then quantise then encode, so the same pipeline is claimed from both ends and again as a bitstream transmission method in claim 7.
• Configurable quantisation in the description. The specification supports dead zone uniform boundary value quantisation and quantisation weighting matrices, and allows the step to be skipped over a whole block or a defined part of one.

Who Is Behind It

The applicant is B1 Institute of Image Technology, Inc., a South Korean company whose published portfolio is video coding and very little else. Its filings run through intra prediction, recursive block division, 360-degree image projection formats and picture scaling for inter prediction, and Ki Baek Kim is named as inventor on nearly all of them. He is the sole inventor on this application. A portfolio of that size attached to a single named engineer is characteristic of a research and licensing business rather than a manufacturer, which matters because codec patents earn their keep through licensing programmes attached to standards rather than through products.

The specification sets out the chain in full. AU 2026202167 is a divisional of Australian application 2024205220, itself a divisional of Australian application 2022204573, itself a divisional of Australian application 2019247240, which was the Australian national phase of PCT/KR2019/003777 filed on 1 April 2019. That PCT application claims the benefit of Korean patent application 10-2018-0037812, filed on 1 April 2018. South Korea is the priority country and the priority date is 1 April 2018, three generations of divisional filings before the present application.

Why It Matters

Of the four applications this family lodged on 19 March 2026, this is the only one whose independent claim describes where the picture is degraded. Its siblings stop earlier. One claims the arithmetic of adding a residual to a prediction. One claims reading residual information out of a bitstream. One claims the inverse transform. This one adds two steps to the body of claim 1 and reaches the quantiser, which is the sole point in the whole chain at which the decoder cannot reproduce what the encoder had. That is a deliberate drafting choice, not a difference in the underlying invention: the description behind all four is identical, and the applicant has simply written the same technique at four different depths of the pipeline and let examination decide how far down a valid claim can reach.

The choice has practical consequences for what such a claim would read on. A decoder that implements a lossless mode, with the quantisation parameter set so the step is bypassed, performs the steps of the shorter sibling claims but does not obviously perform inverse-quantisation at all. Conversely, a claim that names quantisation is anchored to the ordinary operating mode of every consumer video decoder shipping today, because outside of archival and screen-content niches nobody transmits video losslessly. Breadth and grip are traded against each other, and a family with four filings does not have to choose.

The technique itself has a recognisable relative in deployed technology. Splitting an intra coded block into thin strips, predicting them in sequence with one shared mode so that each strip references the reconstruction of the one before it, is the shape of the intra sub-partitions tool in H.266/VVC, published by the ITU in 2020, two years after this family’s Korean priority date. Whether these claims are ultimately held to cover that tool is a question for examiners and, if it ever comes to it, courts. What the filing pattern shows is an applicant treating the question as worth asking four ways at once.

Related Concepts

• Quantization in image processing – the specific operation claim 1 reaches, and the point at which detail is discarded.
• Rate-distortion theory – the framework describing the bitrate against quality trade-off the quantiser controls.
• Quantization in signal processing – the general form of mapping a continuous range of values onto a smaller set.
• Standard-essential patent – the licensing category that gives codec claims their commercial value.
• Divisional patent application – the mechanism that lets one description support four differently scoped filings.

---

AU 2026202167 was published in the Australian Official Journal of Patents on 9 April 2026 and is open for public inspection. Patent applications represent inventions that are sought to be protected and do not necessarily reflect commercially available products.