Learned Image Compression Patents: Who Leads, Where the Gaps Are 2026
- Filing peaked in 2023 at 12 records, then the trend goes flat into the most recent, still-partial year — a maturing rather than an accelerating field.
- One assignee holds 12 of the ranked filings, well ahead of a field where fifth place sits at 4 and tenth place at just 1.
- H04N (pictorial communication) touches 88.9% of the 36 records, while G06V recognition and H01L semiconductor classes each appear in under 6% — a sign of where claims have not yet reached.
What the dataset covers
This landscape draws on 36 published records tagged to learned and neural image compression, spanning claims on rate-distortion optimization, entropy modelling, decoder complexity, perceptual quality metrics and standardization activity, filed under H04N19, G06N3 and G06T9. The scope runs from 2015 through the 2026-07-31 cut-off, though publication lag means the most recent year is undercounted by roughly 18 months relative to actual filing activity.
Twenty-one companies appear in the assignee ranking, headed by a single filer with 12 records against a field where most entrants sit in single digits. The United States is the dominant receiving office, followed by PCT filings through WIPO, then Europe, India, Canada and China.
Let an AI agent run this analysis on your own technology
Pick a task. Every answer cites the patents behind it.
Filing trend and technology composition
Two views of the same 36-record corpus: how filing activity has moved year over year, and which IPC subclasses the claims actually sit in.
A peak in 2023, then flat
Filing rose to a peak of 12 records in 2023 after starting from zero in 2017, then flattened rather than continuing to climb — the 2022 midpoint reads at 0, so the growth curve is back-loaded into 2023 and the immediately surrounding years rather than showing a steady climb across the decade.
Concentrated in pictorial communication and AI computing
H04N (pictorial communication, i.e. video/TV) appears in 88.9% of the 36 records and G06N (AI-model computing) in 55.6% — together these two subclasses cover the bulk of claim activity. G06T (image data processing) trails at 27.8%, while G06V (recognition) and H01L (semiconductor devices) each appear in under 6% of records, marking them as thin rather than saturated ground.
Shares are the percentage of the 36 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on Learned and Neural Image Compression with Eureka
This page is one run against one query. Ask Eureka your own question about learned and neural image compression and every answer comes back with the patent numbers behind it.
Try EurekaThe most-cited filings in scope
Method for Learned Image Compression and Related Autoencoder (US20250088648A1)
The filing describes an autoencoder pipeline: a learnable encoder extracts a latent space from an image, a quantizer discretizes it, an entropy encoder compresses the quantized latent using a probability-distribution entropy model to produce a bitstream, and a matching entropy decoder plus decoder network reconstruct the image. Training is applied end-to-end across the encoder, entropy model and decoder.Filed by Sisvel Technology, published 2025-03-13 — one of the more recent entries in a corpus otherwise weighted toward earlier, more-cited filings.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20200160565A1 | Methods And Apparatuses For Learned Image Compression | 124 |
| 2 | US20200304835A1 | Compression of Images Having Overlapping Fields of View Using Machine-Learned Models | 42 |
| 3 | US20210258611A1 | Compression of Images Having Overlapping Fields of View Using Machine-Learned Models | 6 |
| 4 | US11019364B2 | Compression of images having overlapping fields of view using machine-learned models | 5 |
| 5 | US20250088648A1 | Method for Learned Image Compression and Related Autoencoder | 4 |
| 6 | CN116912520A | 基于空谱特征提取的高光谱端到端压缩方法 | 4 |
| 7 | US20240020887A1 | Conditional variational auto-encoder-based online META-learned image compression | 4 |
| 8 | US11245927B2 | Compression of images having overlapping fields of view using machine-learned models | 3 |
| 9 | WO2024078920A1 | Latent coding for end-to-end image/video compression | 2 |
| 10 | US20240020884A1 | Online meta learning for meta-controlled SR in image and video compression | 2 |
Citation counts are drawn from within this searched corpus and skew toward older filings; treat them as a signal of influence on later filers, not of current commercial weight.
Patent titles are shown in the language they were filed in, not translated, so that each record stays verifiable against the original filing — a translated title will not match in Eureka or in any national register. Each row carries its publication number; clicking a row searches Eureka by that number.
Put your own technology through the same analysis
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →What the numbers mean for a filing decision
Three read-outs from the trend, the IPC mix and the assignee ranking, translated into what they imply for anyone deciding where to file next.
The wave has already crested
Filing climbed to 12 records in 2023 and has not continued upward since; several of the most active named assignees show 0 filings and -100% YoY in the latest year. That pattern is consistent with a field consolidating its core claims rather than one still in an early land-grab phase.
Pictorial communication is the crowded lane
Nearly nine in ten records touch H04N, and over half also carry a G06N AI-computing class, meaning most claims sit at the intersection of video/image transmission and learned models. G06V and H01L barely register, which is where a differentiated claim is more likely to clear prior art.
One filer, then a long tail
The leading assignee holds 12 of the ranked filings; fifth place holds 4 and tenth place just 1, across 21 ranked companies total. That gap suggests a single early mover set much of the claim structure that later, smaller filers had to work around.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to learned and neural image compression, with the prior art for and against each one.
Who is filing, and who is filing together
The ranking spans 21 companies from telecom-standards bodies to automotive and consumer-electronics firms, with a small but tight cluster of co-filers around entropy-model and standardization claims.
A standards-adjacent leader
The top-ranked assignee holds well over double the count of the next tier, consistent with a telecom-standards background where claim portfolios are built deliberately around future codec standardization rather than opportunistic filing.
A tight academic-industry knot
The strongest co-assignee pairs link a technology licensing firm, a French telecom institute and an Italian university research institute, each pair appearing 4-5 times. That triangle looks like a joint-research programme feeding a shared patent strategy rather than independent filers converging by coincidence.
Automotive and platform players sit lower down
Names associated with autonomous-vehicle perception and large consumer-platform AI appear in the ranking but well below the leader, suggesting they are staking claims in learned compression as an adjacent capability rather than a core filing programme.
| Assignee | Recent year | YoY |
|---|---|---|
| InterDigital Madison Patent Holdings | 0 | -100% |
| UATC LLC | 0 | — |
| Sisvel Technology S.r.l. | 0 | -100% |
| Institut Mines-Télécom, Télécom Bretagne | 0 | -100% |
| University of Turin (Università degli Studi di Torino) | 0 | -100% |
| Samsung Electronics Co., Ltd. | 0 | -100% |
| Alibaba DAMO (Hangzhou) Technology Co., Ltd. | 0 | — |
| Xidian University | 0 | — |
Where to take this
The trend and the ranking point to specific next steps depending on whether the goal is freedom-to-operate, portfolio strategy or technical scoping.
Check freedom-to-operate against the leader's claims
With one assignee holding 12 of the ranked filings, any new entropy-model or standardization-adjacent filing should be checked against that portfolio first, not the field average.
Run a claim comparison in Eureka →Scope the under-claimed branches before drafting
G06V and H01L overlaps sit under 6% of records each — thin enough to warrant a targeted prior-art pull before committing claim language there.
Explore white space in Eureka →Watch the co-filing cluster for standardization signals
The densest co-assignee pairs sit around a technology-licensing firm and two research institutes; new standardization proposals are more likely to originate from that cluster than from isolated filers.
Track assignee activity in Eureka →Common questions on learned image compression patents
In this 36-record dataset, one assignee leads with 12 filings, well ahead of the rest of the 21 ranked companies, where fifth place holds only 4 and tenth place just 1. That leader has a telecom-standards background, which is consistent with building a claim portfolio around future codec standardization rather than opportunistic single filings. Anyone assessing freedom-to-operate in this space should treat that portfolio as the first thing to check, not the field average.
Filing activity peaked at 12 records in 2023 and has not continued climbing since, with several of the most active assignees showing 0 filings and a -100% year-over-year change in the latest year on record. That pattern reads as a field settling into its core claims rather than one still expanding rapidly. It is worth remembering that publication lag of roughly 18 months means the most recent year is always undercounted, so a small pickup could still be sitting in the pipeline unpublished.
The dataset is filed under H04N19 (pictorial communication), G06N3 (AI-model computing) and G06T9 (image data processing), and H04N alone touches 88.9% of the 36 records in scope. G06N appears in 55.6% and G06T in 27.8%, with these shares overlapping because most filings carry multiple classes. G06V recognition and H01L semiconductor classes are present but thin, at 5.6% and 2.8% respectively, marking them as the least-occupied adjacent classes in this scope.
US20250088648A1, filed by Sisvel Technology and published 2025-03-13, describes a full autoencoder pipeline for learned image compression: a learnable encoder extracts a latent space, a quantizer discretizes it, an entropy encoder compresses it using a probability-distribution entropy model into a bitstream, and a matching entropy decoder and decoder network reconstruct the image, with end-to-end training across all stages. It sits among the more recent filings in this corpus, in contrast to the highest-cited records which date earlier. Anyone building a similar entropy-model-driven pipeline should read its specific claim boundaries closely rather than assume the general architecture is unclaimed.
The IPC composition points to G06V (image/video recognition) and H01L (semiconductor devices) as the thinnest overlaps, each appearing in under 6% of the 36 records, compared with 88.9% for H04N and 55.6% for G06N. That suggests claims combining entropy modelling with recognition-task conditioning, or with hardware-level implementation, are comparatively under-filed. It is not proof the area is unpatentable, only that the claim space measured here is less occupied than the core H04N/G06N intersection.
Research Learned and Neural Image Compression in depth with Eureka
Go past this page: query the whole learned and neural image compression corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries inherent limitations: recent filings (typically the most recent 18–24 months) are under-counted due to standard publication lag; counts may be reported at either a patent-family or a patent-record basis and are not always directly comparable; classification, applicant-name, and citation data may contain errors, duplicates, or omissions; and the underlying search query defines and constrains the scope shown. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, or validity search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. Any patent, commercial, or strategic decision should be verified independently and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company’s registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.