IBM · AI / ML · Filed May 7, 2025 · USPTO published Nov 12, 2026 · This week's release

IBM Patents Photo Inspection You Set Up With One Example and a Sentence

IBM's new patent application lets you describe a photo-inspection job in a plain sentence, hand over one example picture, and get a defect check without training a new AI from scratch.

By Patentlyze Team · Updated Nov 12, 2026 · 3 min read

A system checks car wheels for manufacturing defects by matching a single sample photo against a brief written inspection instruction. Drawing from patent filing US 2026/0360118 A1. — A system checks car wheels for manufacturing defects by matching a single sample photo against a brief written inspection instruction.All 8 drawings ↓

Publication number: US 2026/0360118 A1(US20260360118A1)
Applicant: INTERNATIONAL BUSINESS MACHINES CORPORATION
Filed · Published: May 7, 2025 · Nov 12, 2026
Inventors: Kun Yan Yin, Wen Wang, Chao Dong, Hanbin Qi, Yi Chen Zhong
Status when we published: Waiting for an examiner
Document: 20 claims
US classification: 382/159
Examiner: VAZ, JANICE EZVI (Art Unit 2667)

AI training & infrastructure

What IBM's one-example photo checker does

You're looking at photos of wheel hubs coming off a factory line, and your job is to catch any with a missing bolt or a scratch. Getting software to do that usually means collecting thousands of labeled photos and having engineers train a special program for that one part.

IBM's patent application describes a shortcut. You hand the system the photo to check, one example photo, and a plain sentence like "Check the quality of the wheel hub. Pay attention to defects." The system breaks that request into smaller jobs, such as finding the right bolts, the missing bolts and the small scratches, and then hunts for each one.

Because it works from a sentence and a sample, the filing says it can take on a new task without retraining, which could cut a lot of setup time.

From the filing · CLAIM 1

… dynamically splitting the visual task into a plurality of sub-tasks by analyzing the text instruction and one or more annotations extracted from the at least one example image; …

Translation: The system breaks down the main job into smaller steps by looking at your written instructions and sample images.

How the request gets split and cross-checked

The request has three parts: the photo to check (the target image), a sentence describing the job, and at least one example image. All three independent claims (a method, a computing device and a software product) describe the same sequence.

First, a large language model (an AI trained to read and write text) studies the sentence and any labels pulled from the example, then splits the job into smaller sub-tasks. In the filing's wheel-hub example, the request "Check the quality of the wheel hub. Pay attention to defects." turns into keywords like "right bolt," "missing bolt," "wrong bolt" and "small scratches."

Then two searches run side by side:

Keyword search: a detector built to find things named in words (the filing mentions one called Grounding-DINO) looks for each keyword in the target photo.
Example search: the example image is copied into several altered versions by rotating, cropping, zooming, or using an image generator (Stable Diffusion) to swap backgrounds or logos. A second detector (the filing mentions DE-ViT) compares those versions with the target photo.

Every detection gets a confidence score. The results are then merged, and a step called non-maximum suppression (when several boxes overlap the same thing, keep only the highest-scoring one) removes duplicates. The output is a labeled image.

From the filing · THE ABSTRACT

The target image is then analyzed separately based on the generated keywords and the augmented images to detect the objects or features in the target image.

Translation: It scans the photo using those keywords and modified examples to spot the items you asked it to find.

Why this matters for factory and scan checks

The practical pitch is less setup. The filing points to factory quality checks and medical image review as places where accurate detection matters but labeled training photos are scarce. If a plant manager could describe a new part in a sentence and supply one good photo, a new inspection job might need far less data preparation.

It also lowers the skill bar. The filing says users should not need technical expertise or a special prompt format, and that the system could run in the cloud or on phones and other local devices. Nothing here names a product, so treat it as a design on paper for now, and remember that a patent application is not proof anything will ship.

IBM's 45th filing we've tracked in our AI training and infrastructure coverage since May follows one where an AI quizzes itself and one penalizing gendered resume words.

Editorial take

The trade here is precision for speed of setup. A system trained on thousands of labeled photos of one product learns that product's quirks. This design learns from one example and a sentence, so it starts out knowing less.

The weak spots are easy to name. The language model has to read a loose sentence and split it into the right sub-tasks, and a bad split means a defect never gets looked for. The invented practice images, such as swapped backgrounds or logos, could also drift away from what real defects look like. The filing gives no accuracy numbers, so the cost of all this flexibility stays unmeasured.

For a factory trying a new part every few weeks, the trade makes sense, because skipping the data-gathering step saves the slowest part of setup. For a hospital reading scans, much more proof would be needed before trusting it.

Source: US 2026/0360118 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time IBM patent news breaks.

Add Patentlyze on Google One click.

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 · 8 sheets

8 drawing sheets from US 2026/0360118 A1 · click any drawing to enlarge

Connected devices sending photos to inspection servers over a network, drawing sheet 1 from US 2026/0360118 A1 — Sheet 1 Connected devices sending photos to inspection servers over a network

Hardware components inside a computer running the inspection software, drawing sheet 2 from US 2026/0360118 A1 — Sheet 2 Hardware components inside a computer running the inspection software

How the system breaks down inspection instructions into simpler steps, drawing sheet 3 from US 2026/0360118 A1 — Sheet 3 How the system breaks down inspection instructions into simpler steps

Steps for finding defects using text descriptions and sample photos, drawing sheet 4 from US 2026/0360118 A1 — Sheet 4 Steps for finding defects using text descriptions and sample photos

Turning one written instruction into specific checklist items to inspect, drawing sheet 5 from US 2026/0360118 A1 — Sheet 5 Turning one written instruction into specific checklist items to inspect

Spotting parts of an object using descriptive text labels, drawing sheet 6 from US 2026/0360118 A1 — Sheet 6 Spotting parts of an object using descriptive text labels

Generating extra training views by cropping, rotating, and altering photos, drawing sheet 7 from US 2026/0360118 A1 — Sheet 7 Generating extra training views by cropping, rotating, and altering photos

Finding and checking parts in a new photo, drawing sheet 8 from US 2026/0360118 A1 — Sheet 8 Finding and checking parts in a new photo

The drawings · 8 sheets

More from IBM

IBM Files Patent for Spotting Tiny Defects in Photos of Factory Parts

IBM Files Patent for an AI That Quizzes Itself and Learns From Expert Corrections

IBM Files Patent to Dock Points From Resume Scores That Lean on Gendered Words

IBM Files Patent to Let Crowds Weed Out Wrong Labels on AI Training Data

More in AI training & infrastructure

Samsung Files Patent for Giving AI Services Only as Much Computing Power as They Need

Microsoft Files Patent for AI That Writes the Blueprint for Fake Test Data

Microsoft Wins a Patent for Training AI on Your Gaming PC's Idle Time

Nvidia Files Patent for AI That Spots and Describes Disease in Medical Scans

No comments yet

Start the discussion Cancel reply