Diffusion Model SAR to EO Image Translation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for translating synthetic aperture radar (SAR) satellite images to electro-optical (EO) satellite images face challenges such as low resolution, sensitivity to electromagnetic characteristics, and speckle noise, which limits their interpretability and accuracy in disaster monitoring.

Innovation Solution

A method using a diffusion model to translate SAR satellite images to EO satellite images by adjusting the diversity of the diffusion model, allowing for the generation of clear and accurate EO satellite images from SAR satellite images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If SAR satellite images are used for disaster monitoring, then it is possible to obtain images through clouds and unaffected by weather, but the images have low resolution and speckle noise that reduces interpretability

Engineering Contradiction:

Improveimage acquisition reliabilityVSAvoidimage resolution

Core Design Contradiction:

ReliabilityVSMeasurement precision

Solution Approach 1:

The patent introduces a diffusion model as an intermediary system that translates SAR satellite images into EO satellite images. The model learns the mapping between SAR and EO image domains, using the cloud-penetrating capability of SAR images while generating output with the visual quality and interpretability of EO images, thus resolving the contradiction between reliability and measurement precision

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent changes the domain parameters by training a diffusion model to transform images from the SAR domain to the EO domain. By adjusting the model's learned parameters and using techniques like test-time augmentation with multiple noise samplings, the system converts low-resolution SAR images into high-resolution EO-like images, improving measurement precision while maintaining the reliability advantage

Inventive Principle:

Principle #35Parameter changes

2Object-affected harmful factors

If denoising technology is applied to SAR satellite images, then speckle noise is removed, but color and object information are not corrected

Engineering Contradiction:

Improvespeckle noiseVSAvoidcolor and object information

Core Design Contradiction:

Object-affected harmful factorsVSLoss of information

Solution Approach 1:

Instead of directly denoising SAR images, the patent uses a diffusion model as an intermediary that translates SAR images to EO images. This translation process inherently removes speckle noise while simultaneously correcting color and object information, as the model learns to generate EO-like images that contain both noise removal and information correction in one step

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent merges the denoising function with the image translation function into a single diffusion model. Rather than applying separate denoising and color correction steps, the model combines these functions, learning to map SAR images to EO images while simultaneously removing noise and correcting information, thus preventing information loss

Inventive Principle:

Principle #5Merging (Combining)

3Adaptability or versatility

If GAN-based image translation is used from SAR to EO satellite images, then translation capability is achieved, but mode collapse problem reduces data quality

Engineering Contradiction:

Improveimage translation capabilityVSAvoidgenerated data quality

Core Design Contradiction:

Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The patent replaces the GAN-based translation mechanism with a diffusion model-based mechanism. Diffusion models work by gradually adding and then removing noise through a learned process, which inherently avoids the mode collapse problem that plagues GANs. This substitution maintains translation capability while significantly improving the quality and diversity of generated images

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent introduces dynamic elements to the translation process by using test-time augmentation with multiple noise samplings. The diffusion model dynamically generates multiple possible translations and aggregates them, which prevents mode collapse by ensuring diverse and high-quality output. This dynamic approach improves manufacturing precision while maintaining adaptability

Inventive Principle:

Principle #15Dynamics

4Adaptability or versatility

If Pix2Pix-based methods are used for SAR to EO image translation, then translation is achieved, but registered images are required for training

Engineering Contradiction:

Improveimage translation functionVSAvoidtraining data requirement

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent inverts the traditional training approach by using unpaired data instead of requiring registered paired images. The diffusion model is trained to translate SAR images to EO images without needing corresponding registered EO images as ground truth, significantly reducing training complexity and data requirements while maintaining translation capability through unsupervised or self-supervised learning strategies

Inventive Principle:

Principle #13The other way round (Inversion)

Data Source

PatentUS20250173834A1Method Of Image-To-Image Translation Using Diffusion Model

Publication Date: 2025.05.29 SI ANALYTICS CO LTD

AI summary

Disclosed is a method for training a diffusion model for image-to-image translation, which is performed by a computing device. The method may include: obtaining an image of a target domain; sampling random noise from a distribution of a source domain; and training a diffusion model that translates an image of the source domain to the image of the target domain based on the sampled noise.