Turning a Prediction Error Into Frequencies Why This Divisional Claims the Transform Step

This application claims a way of compressing detail inside a single video frame. A square block is cut into four equal squares by a tree, one of those squares is then sliced by an index into four thin strips or two halves, and every piece is predicted using the same intra prediction mode. What separates this filing from three siblings lodged on the same day by the same applicant is one clause near the end of claim 1: the leftover error signal is recovered by inverse transforming information on the residual block. It was filed by B1 Institute of Image Technology, Inc. of South Korea, with Ki Baek Kim named as sole inventor.

The Problem

A codec does not store pixels. It guesses them, then stores the size of its mistake. For a block that has no earlier frame to borrow from, the guess has to be built out of pixels already reconstructed along the top and left edges of the block itself, extended inwards along a chosen angle or averaged flat. That is intra prediction, and the specification’s stated objects are narrow: to provide a method and apparatus for intra prediction, for subblock-wise intra prediction, and for partitioning into subblocks and determining a coding order of those subblocks.

The weakness of predicting a whole block at once is distance. A prediction built from the row of pixels above a 32 by 32 block is a reasonable estimate for the first row inside it and a poor one for the thirty-second, because the further the fill travels from its reference edge the less the reference has to say. The specification puts the point concretely. When encoding is performed at the parent block level, the bottom-right region has to be predicted from pixels neighbouring the parent block. When the same area is encoded as a subblock, it can be predicted from closer pixels, because the top-left, top-right and bottom-left subblocks have already been reconstructed in a set coding order and are available as references.

Splitting is not free. Every extra block boundary normally costs a fresh set of signalled parameters, and a codec that pays full price for each fragment gives back in overhead what it gained in accuracy. The specification’s answer, and the one written into claim 1, is that the pieces share. One prediction mode is generated and used for every subblock, so the split buys shorter reference distances without buying four times the mode signalling. As the specification says, since only one prediction mode is generated and shared between the subblocks, the subblock-wise encoding may be efficient.

The remaining problem is what to do with the error that survives all of that. Even a good prediction leaves a residual, and a residual stored as raw differences is barely smaller than the picture it came from.

What This Invention Does

Claim 1 is a decoding method and it runs in a fixed order. First it generates a coding block using tree-based partitioning, which the claim defines tightly as splitting a square block into four square blocks of the same size, in other words a quad split and nothing else. Then it generates a plurality of sub-blocks by performing index-based partitioning on that coding block.

That second split is the unusual part. The claim gives it two named operations. A first partitioning takes a 4Nx4M coding block and produces four identical 4NxM sub-blocks or four identical Nx4M sub-blocks, with N equal to or different from M, which is the thin-strip case. A second partitioning simply splits the coding block into two equal parts horizontally or vertically. Which of the two applies is not signalled freely: a closing wherein clause states that whether to perform the first partitioning or the second partitioning is determined based on the size of the coding block. Small blocks would produce uselessly thin strips under a four-way split, so the size test keeps the strip mode away from them.

The specification separates this from the tree splits deliberately. In tree-based partitioning a partition type is determined from several pieces of information, such as whether partitioning is performed, a tree type and a direction. In index-based partitioning a type is chosen with a single index into a candidate group, which the specification says suits single-layer splitting rather than recursive splitting, and a sub-block produced this way is not split any further.

Intra prediction then runs on each sub-block, and another wherein clause requires that it be performed based on the same intra prediction mode for every one of them. A residual block is generated, and the coding block is reconstructed from the prediction block and the residual block.

The distinguishing limitation sits between those last two steps. In AU 2026202166 the residual block is generated by inverse-transforming for information on the residual block. The matching encoder claim, claim 5, states that the coding block is encoded by performing transformation on the residual block. Claim 7 wraps the same encoder chain in a bitstream transmission method. Nothing in claim 1 mentions quantisation, and nothing in it mentions parsing a bitstream. The claim reaches exactly as far as the transform and stops.

Key Features

• A quad split defined in the claim itself. Tree-based partitioning is confined to splitting a square into four equal squares, so the claim does not depend on binary or ternary tree variants to establish the coding block.
• Two index-based partitionings, chosen by size. The first cuts a 4Nx4M block into four identical strips in one direction, the second halves the block, and the coding block’s size decides which is available.
• One intra prediction mode across all sub-blocks. Every sub-block is predicted with the same mode, which is what keeps the signalling cost of the split down while shortening the distance from each sub-block to its reference pixels.
• Index-based splitting is single-layer. The specification states that a sub-block obtained through index-based partitioning is not split again, either into further coding blocks or into separate prediction blocks, so the structure stays shallow.
• Residual recovered by inverse transformation. The decoding claim recites that the residual block is generated by inverse-transforming information on the residual block, which is the limitation that distinguishes this application from its same-day siblings.
• A transform block sized to the prediction block. The specification’s worked examples include configurations where the transform block takes the size of the prediction block obtained by the split, so the strips are transformed as strips rather than as one parent block.

Who Is Behind It

The applicant is B1 Institute of Image Technology, Inc., a South Korean company whose patent portfolio is almost entirely video coding. Its filings, visible in bulk on public patent databases, cover intra prediction, block partitioning, 360-degree image projection formats and inter prediction scaling, and they name the same person as inventor almost every time. Ki Baek Kim is the sole inventor here and the sole or lead inventor across the great majority of the company’s published applications, which is an unusual concentration for a codec portfolio and suggests a research-and-licensing operation built around one engineer’s work rather than a product company.

The chain of filings is stated plainly in the specification and is unusually long. AU 2026202166 is a divisional of Australian application 2024205220, which was itself a divisional of Australian application 2022204573, which was itself a divisional of Australian application 2019247240. That application was the Australian national phase of PCT/KR2019/003777, filed on 1 April 2019 and published as WO 2019/194485, which claims the benefit of Korean patent application 10-2018-0037812 filed on 1 April 2018. South Korea is therefore the priority country, and the priority date is eight years before this filing.

Why It Matters

The transform is the step that makes a residual worth storing. A block of prediction errors in the spatial domain is a grid of small numbers scattered more or less everywhere. Run a discrete cosine transform over it and the same information becomes a grid of coefficients in which a few low-frequency entries near one corner carry most of the energy and the rest hover near zero. Nothing has been discarded yet, but the data has been rearranged into a shape where discarding is cheap and where the near-zero majority costs almost nothing to encode. The specification describes exactly this move, transforming a spatial signal to a frequency signal, and lists Hadamard, discrete sine and discrete cosine schemes as options, with sub-variants DCT-1 to DCT-8 and DST-1 to DST-8 and candidate groups such as DCT-2, DCT-8 and DST-7. That trio is the same set adopted for multiple transform selection in H.266/VVC.

Reciting the transform is a real narrowing rather than a formality, because the specification is explicit that the step is optional. Some transformation may be skipped according to the encoding configuration, it says, and a residual block may be subjected only to quantisation without transformation. A claim that requires inverse transformation therefore covers a smaller set of codec behaviours than one that merely requires a residual to be produced.

That is the logic of the family. Four applications were filed on 19 March 2026 with an identical core, and they differ only in how far down the coding pipeline the independent claim reaches: subtraction and addition in one, the bitstream in another, this one at the transform, and a fourth going all the way through quantisation. Australian practice allows a divisional to be filed off a pending parent, so an applicant with a long-lived family can keep prosecuting a ladder of claim shapes and let examination decide which rungs hold. For a technique that resembles the intra sub-partitions tool in a deployed standard, that ladder is the asset. The strip-wise prediction described here is close in spirit to what was standardised in H.266, and the value of a claim reading on a shipped codec depends far less on the elegance of the idea than on which layer of the pipeline the surviving claim happens to sit at. Compared with its siblings, this one sits at the layer where the residual stops being a picture and becomes a spectrum.

Related Concepts

• Coding tree unit – the top-level block that the quad split in claim 1 works down from.
• Quadtree – the data structure behind splitting a square into four equal squares.
• Discrete sine transform – the DST family the specification lists alongside DCT as a transform scheme for the residual.
• Joint Video Experts Team – the standards group that developed the H.266 tools this technique parallels.
• High Efficiency Video Coding – the earlier standard whose block partitioning vocabulary the specification builds on.

---

AU 2026202166 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.