IBM · Filed Jan 28, 2025 · Published Jul 30, 2026 · verified — real USPTO data

IBM Patents an AI Method for Sharpening Low-Resolution Images

Blurry images are an old problem, but IBM is tackling it with an AI approach that teaches itself which visual patterns matter most before trying to fill in the missing detail.

IBM Patent: High-Resolution Image Reconstruction — figure from US 2026/0220739 A1
Figure from the official USPTO publication.
See all 9 drawings from this filing ↓
Publication number US 2026/0220739 A1
Applicant INTERNATIONAL BUSINESS MACHINES CORPORATION
Filing date Jan 28, 2025
Publication date Jul 30, 2026
Inventors Kyong Min YEO, Takuya KURIHANA, Levente KLEIN, Daniela DE MATTOS SZWARCMAN
CPC classification 382/156
Grant likelihood Medium
Examiner CENTRAL, DOCKET (Art Unit OPAP)
Status Docketed New Case - Ready for Examination (Feb 26, 2025)
Document 20 claims

What IBM's image-sharpening AI actually does

Imagine you have a photo taken from a weather satellite. The image is fuzzy because the camera only captured a low number of pixels, and zooming in just makes it blurrier. What IBM's system does is take that blurry input and try to reconstruct a sharper, more detailed version automatically.

The AI doesn't just guess at the missing pixels. It first studies the structure of the blurry image to understand how information is arranged spatially, for instance how temperature or color changes from one region to the next. It then builds a kind of attention map, a picture of which parts of the image carry the strongest, most reliable signals.

Finally, it applies what the patent calls a "penalty term," which turns up the volume on the strong signals and turns it down on the weak, noisy ones. The result is a high-resolution reconstruction that preserves the real patterns rather than inventing details that were never there.

How the attention map amplifies the right signals

The method starts by running the low-resolution image through a two-dimensional convolutional filter (a standard AI technique that slides a small detection window across an image to spot local patterns like edges or gradients). From that, the system builds an internal representation of the image's spatial structure.

That representation is then replicated into multiple versions, each aligned with different mathematical vectors. The system computes a self-attention map by taking the dot product between these representations (dot product here means multiplying two sets of numbers together and summing them, a way to measure how similar or related two things are). The attention map captures what the patent calls "vertical mixing of information," meaning how values at one spatial location relate to values at other locations, even if they are far apart.

The final step applies a penalty term to that attention map:

  • Strong, consistent signals get amplified
  • Weak or noisy signals get suppressed
  • The result feeds into the final high-resolution output

This penalty step is essentially the system's quality filter: it prevents the model from confidently reconstructing details that were never reliably present in the original data.

What this means for satellite and medical imaging

The most immediate applications are fields where low-resolution data is a hard physical constraint, not just a camera limitation. Satellite imagery, weather modeling, and medical imaging (like low-dose CT scans) all involve trade-offs between resolution and cost or safety. A method that reconstructs finer detail without introducing hallucinated artifacts could meaningfully improve how scientists and clinicians interpret that data.

For IBM, this fits into a broader push around geospatial AI, an area where the company has been active with foundation models trained on satellite and climate data. If this technique is integrated into those pipelines, it could sharpen the inputs those models see, which matters because better input data generally leads to better predictions.

Editorial take

This is solid, specific AI research rather than a flashy consumer-facing idea. The penalty-term mechanism for sharpening attention maps is the genuinely interesting piece here: it addresses a real failure mode in image super-resolution where models confidently invent plausible-looking but wrong detail. Whether IBM turns this into a product or it stays in the research layer, the technique is worth watching in satellite and climate-data contexts.

The drawings

9 drawing sheets from US 2026/0220739 A1 · click any drawing to enlarge

Patent filing page

Which company should we read for you?

We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.

Get one Big Tech patent every Sunday

Plain English, intelligent commentary, no hype. Free.

Source. Full patent text and figures from the official USPTO publication PDF.

Editorial commentary on a publicly published patent application. Not legal advice.