Four Strips and One Prediction Mode The Video Coding Claim That Only Adds and Subtracts

This application claims a way of cutting a square block of a video frame into four thin strips, predicting each strip separately but with a single shared prediction rule, and then rebuilding the block by adding the prediction to whatever the prediction got wrong. What makes it unusual is what claim 1 leaves out. There is no bitstream in it, no transform, no quantisation, nothing about how any of this travels or is stored. It was filed by B1 Institute of Image Technology, Inc., a Korean video coding research company, and names Ki Baek Kim as sole inventor.

The Problem

A video frame is not compressed all at once. It is carved into blocks, and each block is guessed at from pixels that have already been decoded around it, usually the row above and the column to the left. That guess is the prediction block. The difference between the guess and the truth is the residual, and because a good guess leaves a small residual, the whole business of intra frame coding is a competition to make the guess better so there is less left over to send.

The specification’s own framing is broad. It notes that multimedia use has grown with the internet and portable terminals, that image processing systems need better performance and efficiency, and that research has not kept pace. Its stated objectives are narrower and more useful: to provide intra prediction performed on a subblock basis, and to provide a way of partitioning a block into subblocks and choosing the order they are coded in.

The reason subblocks matter is a geometry problem. Reference pixels sit along the top and left edge of a block. If you predict a large square as one unit, the pixels in its bottom right corner are as far as they can possibly be from any reference pixel, and directional prediction degrades over that distance. The specification says this plainly. When encoding happens at the parent block level, the bottom right region has to be predicted from pixels neighbouring the whole parent block. When it happens at subblock level, that same region can be predicted from much closer pixels, because the top left, top right and bottom left subblocks have already been reconstructed and are now available as references themselves.

The catch is cost. Splitting a block normally means signalling a new prediction mode for every piece, and that overhead can wipe out the gain. So the interesting question is not whether to split, but how to split without paying for it four times over.

What This Invention Does

Claim 1 is an image decoding method, and it runs in five steps.

First, a coding block is generated using tree based partitioning, which the claim defines tightly as splitting a square block into four square blocks of the same size. That is the ordinary quad split familiar from the quadtree structures used to carve up a coding tree unit in modern codecs.

Second, that coding block is split again, this time by index based partitioning rather than by a tree. The distinction matters and the specification spells it out: in tree based partitioning the shape is worked out from several pieces of information such as whether to split, which tree type and which direction, whereas in index based partitioning the shape is picked in one shot using an index. Claim 1 defines two flavours. The first partitioning takes a 4N by 4M coding block and splits it into four identical 4N by M sub-blocks, or four identical N by 4M sub-blocks. In other words, four flat horizontal strips or four narrow vertical ones. The second partitioning simply halves the block, horizontally or vertically.

Third, a prediction block is generated by performing intra prediction on each of those sub-blocks. Crucially, the claim requires that every sub-block use the same intra prediction mode. One mode is derived for the parent and shared by all its strips, so the split buys shorter reference distances without buying four sets of mode information.

Fourth, a residual block is generated. Fifth, the coding block is reconstructed from the prediction block and the residual block.

Then come the wherein clauses, and this is where AU 2026202164 parts company with its three siblings. It states that the coding block is reconstructed by adding the prediction block and the residual block. The matching encoder claim, claim 5, states that the residual block is generated by subtracting the prediction block from the coding block. That is the entire arithmetic of a predictive codec, stated in two words, and nothing else about the pipeline appears in claim 1 at all.

The last wherein clause is the size gate. Whether the first partitioning or the second partitioning applies is determined by the size of the coding block. The body explains why that is not arbitrary. A codec imposes minimum block dimensions, so a four way strip split may simply be unobtainable for a small block, or for a block sitting on a picture boundary. In that case, the specification describes falling back through a priority order of partition types, from four strips to two halves, and if none of the predefined types is available, leaving the parent block unsplit. Reading the claim against the body, the size condition is the rule that decides whether there is room for four strips or only for two.

Key Features

• Two stage partitioning. A square is first quad split into four equal squares by the tree, and only then split a second time by index into strips. The strips are a single non recursive operation, not another level of tree.
• Four strips instead of four squares. The first partitioning produces four identical 4N by M or N by 4M sub-blocks, deliberately long and thin, so that each strip is close to the reference pixels of the one before it.
• One intra prediction mode for all sub-blocks. Every sub-block is predicted using the same mode, derived once at the parent level, which is what stops the split from multiplying the side information.
• A coding order that shortens the reference distance. The specification ties scan order to split direction, using a vertical scan for horizontally split parents and a horizontal scan for vertically split ones, so each strip is reconstructed just in time to serve as reference for the next.
• A size gate between the two partitionings. Which of the two index based partitionings applies is decided by the size of the coding block, with the body describing minimum width and area conditions that make the four way split unavailable for small or boundary blocks.
• Reconstruction stated as pure addition. Claim 1 recites only that the block is reconstructed by adding prediction and residual, and claim 5 only that the residual is produced by subtraction, with no transform, quantisation or bitstream anywhere in the independent claims.

Who Is Behind It

The applicant is B1 Institute of Image Technology, Inc., a Korean research company whose output is almost entirely video coding intellectual property rather than products. It appears on the licensor list for the HEVC and VVC patent portfolio licence now administered by Video Codec Licensing, a subsidiary of Access Advance, which places it among the declared essential patent holders for the standards this technology sits alongside.

Ki Baek Kim is the sole named inventor, and is the sole or lead inventor across a large run of B1 filings covering intra prediction, block partitioning, projection formats for 360 degree video and related subjects.

The Australian chain here is long, and the specification sets it out in full at the head of the body. AU 2026202164 is a divisional of Australian application 2024205220, itself a divisional of 2022204573, itself a divisional of 2019247240, which is the Australian national phase of an international application filed on 1 April 2019 claiming the benefit of Korean patent application 10-2018-0037812 filed on 1 April 2018. The specification prints the international application number as PCT/K2019/003777, evidently a typographical slip; the published family record shows it as PCT/KR2019/003777, published as WO 2019/194485. That correction comes from the published record, not from the Australian specification, which is otherwise explicit about the whole chain.

Why It Matters

The technique in this family is close in spirit to the Intra Sub-Partitions tool standardised in Versatile Video Coding, the H.266 successor to High Efficiency Video Coding. That tool divides a block into vertical or horizontal sub-partitions predicted in sequence under one shared mode, for the same reason described here: prediction quality falls off with distance from the reference row, and processing strip by strip keeps the distance short. The priority date of 1 April 2018 sits in the middle of the standardisation activity that produced it.

The more interesting thing about this particular document is that it is one of four applications filed on the same day, by the same applicant, with the same inventor, the same title, the same 133 pages of description and an almost word for word identical claim 1. The shared part of the claim is the whole story above: quad split, index based partitioning into strips or halves, one shared intra mode, size gated choice between the two. The only text that differs is the clause describing how the residual comes to exist, and the four applications place that clause at four successive layers of the codec.

This one sits at the bottom. Its claim describes the residual purely as an arithmetic relationship between two blocks of pixels: subtract to make it, add to undo it. Nothing about frequency domain representation, nothing about how coarsely values are rounded, nothing about how any of it is written down. A claim drawn this way is about the shape of the block and the order of operations, so an implementation that skipped the transform entirely, or that coded losslessly, would still be doing every step the claim recites. Its siblings each add one more layer, and each of those layers narrows the claim while making it easier to point at in a real file or a real chip. Filing all four is a way of not having to choose which trade off is the right one until much later, when the prior art has been tested and it is clear which rung of the ladder can actually be defended.

Related Concepts

• Image compression – the broader field the block prediction and residual model belongs to.
• Lossy compression – the trade off that makes a small residual worth chasing in the first place.
• Macroblock – the fixed size ancestor of the flexible coding block this claim partitions.
• Video Coding Experts Group – one of the two standards bodies whose joint work this technology was developed alongside.
• Divisional patent application – the mechanism that lets one description support four differently scoped claims.

---

AU 2026202164 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.