Samsung Patents a Way to Pick the Right Decoding Math for Each Video Block
Every video you watch is secretly a pile of math puzzles, and Samsung thinks the decoder has been solving some of them with the wrong tool. This patent describes a method that lets the decoder pick the right formula for each chunk of a frame, based on what the picture actually looks like.
What Samsung's per-block transform selection actually does
Video files don't store pictures the way a camera captures them. Instead, they store instructions for rebuilding each tiny block of each frame, and those instructions have to be "decoded" using math. Today's decoders often apply a one-size-fits-all formula, even when a particular block of pixels would reconstruct more accurately with a different approach.
Samsung's patent describes a decoder that looks at a block's reference block (the nearby pixels used to predict what the current block should look like) and figures out the local texture or direction of detail in that spot. From that analysis, it picks the specific math formula, called a transform kernel, best suited to that region of the image.
The practical goal is cleaner reconstruction, especially in areas with sharp edges, diagonal patterns, or fine textures, which are exactly the places where standard decoding tends to blur or ring. It's a targeted fix rather than a blanket upgrade.
… determining a transform kernel of a current block from among one or more transform kernels by using the intra mode for the reference block; and performing inverse transformation on the current block by using the transform kernel.
Translation: Choosing the right math formula to decode a video block based on how a nearby block was set up.
How the decoder reads neighbors to choose a transform kernel
The patent covers both an encoding side and a decoding side, but the core idea sits in the decoder. When reconstructing a block of a video frame, the system first retrieves a reference block from a nearby or previously decoded region of the image.
Next, it analyzes the reference block using its neighboring samples (pixels around the edges of the block) or its internal samples (pixels inside the block) to determine an intra prediction mode (a label, essentially, that describes which direction the texture or edge flows: horizontal, vertical, diagonal, and so on). This is the same kind of directional analysis that modern video codecs like H.266/VVC already use during encoding, but here it's being re-derived at decode time to inform a separate choice.
That choice is the transform kernel: the specific mathematical function used to convert the compressed difference data back into pixel values. Standard decoders typically default to a Discrete Cosine Transform (DCT), which works well on smooth gradients but can struggle with sharp edges. By mapping intra modes to better-matched kernels (such as a Discrete Sine Transform or a hybrid variant), the decoder can recover more detail in tricky regions.
- Obtain the reference block from the reference image
- Derive the intra prediction mode from neighboring or internal samples
- Select the transform kernel that matches that mode
- Apply the inverse transform to reconstruct the current block
What this means for video quality on Samsung devices
For most viewers, better transform selection means fewer visible artifacts around sharp edges and fine textures in compressed video, the kind of ringing or smearing that shows up in fast-action sports footage or high-contrast graphics. The gain would come without needing a higher bitrate, which matters for streaming over limited connections.
The method sits squarely in the territory of next-generation video codec design, where Samsung has significant commercial interest across televisions, phones, and chipsets that handle video decoding. Whether this ends up in a future codec standard or in a proprietary Samsung implementation is an open question, but the target is clearly the layer of the stack that every video-playing device touches.
Samsung's 41st filing in the wearables patents we cover since May follows work like the earbud audio sync fix and the photo-driven watch face.
The cost Samsung is paying here is computation on the viewer's side. Instead of storing a small piece of information in the video file, the design makes your phone or streaming device re-calculate that information during playback. That shifts work away from the file and onto the hardware in your hands.
Whether that exchange holds up depends on how fast the re-calculation runs. If it runs quickly, the video file gets smaller and your device never breaks a sweat. If it runs slowly, your battery drains faster and the picture quality improvement is probably too small to see, which makes the trade a loss.
Samsung has identified a real problem with how compressed video chooses its math operations, and the approach is methodical. The honest gap is that this document includes no measurements, so the trade either pays off brilliantly or fails in the implementation, and nothing here tells you which.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
38 drawing sheets from US 2026/0281399 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →