Machine Learning Image Sharpening via Synthetic Blur Training

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning-based technologies face challenges in creating effective training data for image processing tasks, such as super-resolution, and struggle to balance noise reduction and image sharpening in mobile device cameras.

Innovation Solution

A machine learning-based approach that generates synthetic training data by introducing controlled blur to ground truth images using demosaic and noise filtering operations, and uses a trained model with shallow and deep feature extractors to sharpen images by removing processing blur.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If noise filtering is applied to reduce noise in images, then noise reduction is improved, but image sharpness deteriorates due to blurry edges

Engineering Contradiction:

Improvenoise reductionVSAvoidimage sharpness

Core Design Contradiction:

ReliabilityVSManufacturing precision

Solution Approach 1:

The patent applies inversion by training the machine learning model to reverse the blurring effect caused by noise filtering. Instead of accepting the blurred output of traditional noise filtering, the system inverts the process by learning the mapping from blurred images to sharp images, thereby recovering edge sharpness while maintaining noise reduction benefits.

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The patent converts the harmful blur introduced by noise filtering into a beneficial training signal. By using the blurred images as training data and teaching the model to map them to ground truth sharp images, the system transforms the harmful effect of blur into a useful learning opportunity that improves the model's ability to restore sharpness in real applications.

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

2Quantity of substance

If high-resolution images are downsampled to create training data for super-resolution, then training data availability is improved, but the model learns upsampling instead of super-resolution

Engineering Contradiction:

Improvetraining data availabilityVSAvoidmodel functionality

Core Design Contradiction:

Quantity of substanceVSManufacturing precision

Solution Approach 1:

The patent applies preliminary action by pre-processing the training data to include realistic blur and noise effects before training the model. This ensures that the model learns to handle degraded images properly and performs true super-resolution rather than simple upsampling, while still using the abundant data available from downsampling high-resolution images.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12272032B2Machine learning-based approaches for synthetic training data generation and image sharpening

Publication Date: 2025.04.08 SAMSUNG ELECTRONICS CO LTD

AI summary

A method includes obtaining an input image that contains blur. The method also includes providing the input image to a trained machine learning model, where the trained machine learning model includes (i) a shallow feature extractor configured to extract one or more feature maps from the input image and (ii) a deep feature extractor configured to extract deep features from the one or more feature maps. The method further includes using the trained machine learning model to generate a sharpened output image. The trained machine learning model is trained using ground truth training images and input training images, where the input training images include versions of the ground truth training images with blur created using demosaic and noise filtering operations.