The assignee ranking in this dataset covers 100 companies across 116,090 published records, with the leading assignee holding 16,204 records — roughly three times the fifth-place total of 5,331. The five leading assignees combined account for 41.6% of all records in scope, so the field is concentrated at the top even though the full ranked list runs to 100 companies. Anyone assessing competitive position should look at family-level counts rather than raw document counts, since a single filing can generate multiple published documents.

Filings grew from 1,798 records in 2021 to 2,226 in 2024, a +24% increase, and 2024 is the most recent year the trend can treat as complete. Later years in the dataset show lower counts, but that reflects the roughly 18-month lag between filing and publication rather than a real slowdown. The honest read is that activity was still rising through the last complete year, with 2020 as the peak year so far at 2,951 filings.

H04N, the pictorial communication and video/TV subclass, dominates at 17.0% of all 116,090 records — far ahead of every other subclass. G06T, H03M, G06K, G06N, H04B, G06F and G06V each account for 0.4% of records or less. Because a single record can carry multiple IPC classes, these shares add up to more than 100%, and any class-level share should always be read against the full record total, not against another class's count.

US20260095600A1 is a Samsung filing published 2026-04-02 that claims an AI-based in-loop filter selecting a pre-processor from multiple candidates based on features extracted per frame and channel. It matters because it frames filter selection as a learned, feature-conditioned step rather than a fixed rule, which narrows the room for a near-identical adaptive-selection claim. Anyone drafting in-loop filter claims that involve any form of feature-based pre-processor selection should review its specific claim language before finalising scope.

The IPC composition shows core video-communication art (H04N) is heavily claimed while adjacent AI-model and image-recognition subclasses (G06N, G06V) sit at 0.1% of records each, suggesting thinner coverage where codec tools intersect with learned models. Sub-areas such as AI-conditioned in-loop filter selection, learned quantization-parameter adaptation, and entropy coding tuned for AI-generated residuals show comparatively low filing density relative to the core toolset. These are not guaranteed clear, but they are the branches where a first claim has more room to stand without colliding with dense prior art from the leading assignees.