Semantic Map Image Blending for Authentic Satellite Manipulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing generative adversarial networks (GANs) struggle with blending artifacts when generating images outside their training domain, particularly in geographical map-to-satellite image transformations, leading to perceptible changes in manipulated images.
Innovation Solution
An image generation system uses a generative neural network (GAN) to modify semantic map images, blending GAN-generated satellite images with pristine satellite images using techniques like Poisson Image Editing to create deepfake satellite images that preserve original pixels and authenticity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If GAN-generated satellite images are blended with pristine satellite images using Poisson Image Editing, then image quality and authenticity are improved, but device complexity and processing difficulty increase
Solution Approach 1:
The patent uses Poisson Image Editing as an intermediary technique to blend GAN-generated satellite images with pristine satellite images. This mediator process harmonizes the two image sources by solving the Poisson equation to match gradients and intensities, thereby reducing visible artifacts and improving overall image quality while maintaining a systematic approach to the complex blending task
Solution Approach 2:
The patent segments the image manipulation process into distinct stages: (1) generating manipulated semantic map images with desired changes, (2) using GANs to translate these maps to satellite imagery, and (3) blending the generated images with pristine images using Poisson editing. This segmentation allows each stage to be optimized independently, managing complexity through structured decomposition
2Adaptability or versatility
If GANs are used to generate images outside their training domain, then versatility and application range are improved, but blending artifacts and image quality deteriorate
Solution Approach 1:
Poisson Image Editing serves as a mediator that corrects the artifacts introduced when GANs generate images outside their training domain. By using the pristine image as a reference and blending the generated image with it through gradient matching, the technique reduces domain mismatch artifacts while preserving the versatility of generating images for various applications including urban planning, agriculture, and wildfire analysis
Solution Approach 2:
The patent applies local quality adjustments by using Poisson blending specifically in regions where GAN-generated images show artifacts or domain mismatch. The blending operation is applied selectively to harmonize local regions by matching gradients and intensities with the pristine image, rather than uniformly processing the entire image, thereby maintaining versatility while improving local image quality
Data Source
AI summary
Systems and methods herein describe an image generation system that accesses a semantic map and satellite image, manipulates the semantic map image, trains a machine learning framework using a set of map and satellite image pairs, uses the trained machine learning framework to generate a manipulated satellite image based on the manipulated semantic map, generates a blended satellite image based on a combination of the manipulated satellite image data and the originally accessed satellite image data, and stores the blended satellite image.