Samsung
Samsung · AI / ML · Filed Oct 15, 2025 · USPTO published Jun 25, 2026 · Jun 25 release
Samsung Patents a Way to Train Video AI on Deliberately Distorted Frames
Samsung's latest patent describes a training trick for video AI: deliberately mess up a frame, then punish the model if its answers change. The idea is that a truly good AI should recognize the same objects whether the image is clean or slightly scrambled.
- Publication number
- US 2026/0179382 A1(US20260179382A1)
- Applicant
- Samsung Electronics Co., Ltd.
- Filed · Published
- Oct 15, 2025 · Jun 25, 2026
- Inventors
- Mostafa EL-KHAMY, Kareem METWALY
- Status when we published
- Waiting for an examiner
- Document
- 20 claims
- US classification
- 382/155
- Parent application
- Claims priority from a provisional application 63738576 (filed Dec 24, 2024)
What Samsung's jittered-frame video training actually does
Imagine you're learning to recognize faces. A good teacher might show you a photo, then show you the same photo with the brightness slightly off or the image a little blurry, and ask: 'Is that still the same person?' If your answer changes, you need more practice.
Samsung's patent applies that same logic to video AI. The system takes a single video frame, creates a slightly distorted copy of it (think: minor warping, color shift, or blur), and then checks whether the AI's identification of objects stays the same across both versions. If the AI's answer flips, the training process penalizes it.
The goal is a video AI that labels objects consistently from one moment to the next, even when lighting or camera motion causes each frame to look a little different. That kind of stability is hard to get without labeling thousands of hours of footage by hand, which is exactly what this approach tries to avoid.
How the distortion-comparison loop trains the network
The patent describes a semi-supervised training method for video segmentation, meaning the AI can learn from mostly unlabeled video data rather than requiring every frame to be hand-annotated by a human.
The core step is temporal regularization (a penalty applied during training that pushes the model toward consistent predictions over time). Instead of comparing the current frame to the previous frame in the video, the system generates a jittered version of the same frame by applying a computer-controlled image distortion. The model then produces segmentation predictions (pixel-by-pixel labels like 'this is a car,' 'this is a person') for both the original and the jittered copy.
- If the predictions agree, the model gets a good score.
- If the predictions diverge, a loss penalty is applied and the model's weights are adjusted.
- This loop repeats across training data, nudging the network toward stability.
The result is a frame-level segmentation network that has learned to ignore small visual inconsistencies, the kind that show up constantly in real video due to motion blur, compression, or lighting changes.
What this means for AI video editing and recognition
Video segmentation is the backbone of a wide range of features: background replacement in video calls, object tracking in cameras, scene understanding for autonomous systems, and AI-powered video editing. Getting it to work consistently across frames without expensive labeled data is an ongoing challenge for the whole industry.
For Samsung specifically, this kind of technique could find its way into on-device camera processing, Galaxy AI video tools, or computer vision systems in their broader electronics lineup. The semi-supervised angle is the practical win here: if you can train a reliable model with far fewer hand-labeled frames, you cut the cost and time of building video AI products significantly.
Source: US 2026/0179382 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 Samsung 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.
More from Samsung
All 1,629 Samsung patents →
More in AI vision
All 339 AI vision patents →
IBM
IBM Patent: Tracing AI Reasoning Paths to Detect Corrupted Trained Models
Google
Google Patents a Way to Teach AI When to Say 'The Answer Isn't Here'
Microsoft
Microsoft Patent Targets Fraud Patterns Slipping Past AI Detection Models
IBM
