Semantic Image Extrapolation With Segmentation-Guided Inpainting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image editing and compositing technologies struggle to create natural composite images by simply enlarging or reducing images without considering the ratio of object to background, leading to unnatural results.
Innovation Solution
A semantic image extrapolation method using an extrapolated segmentation map and inpainting technique to fill empty regions in images, generating a padded image and combining it with the extrapolated segmentation map to create a new image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Length of moving object
If the entire image is simply enlarged or reduced, then the image size is changed, but the composition becomes unnatural due to not considering the ratio of object image to background image
Solution Approach 1:
The patent divides the image into object regions and background regions using segmentation maps. This allows independent processing of object and background, enabling selective enlargement or reduction of specific regions while maintaining natural composition ratios. The segmentation map identifies which pixels belong to objects versus background, permitting differential scaling that preserves compositional naturalness.
Solution Approach 2:
The patent applies different transformation qualities to different regions of the image. Object regions are processed differently from background regions, with the system adjusting the degree of enlargement or reduction based on the specific characteristics and ratios of each region. This local differentiation ensures that the composition remains natural while achieving the desired overall image size change.
2Ease of manufacture
If image editing/compositing technology is used to erase or separate image parts, then desired parts can be extracted, but it becomes difficult to provide a natural composite image when extending images
Solution Approach 1:
The patent generates segmentation maps and extrapolated segmentation maps in advance before performing image extension. These pre-computed maps provide guidance for how to extend the image while maintaining natural composition, enabling the system to handle extension tasks that would otherwise be difficult or produce unnatural results.
Solution Approach 2:
The patent introduces segmentation maps and extrapolated segmentation maps as intermediary structures between the original image and the extended image. These intermediaries contain spatial and compositional information that guides the extension process, allowing the system to seamlessly integrate new regions while preserving the natural appearance and composition of the original image.
Data Source
AI summary
Disclosed are a semantic image extrapolation method and a semantic image extrapolation apparatus. The present invention provides a technique for generating an empty region for image-extension in an image by using an extrapolated segmentation map and an inpainting technique. The present invention is to provide, considering that there is no information in an empty region for image-extension in an image, a semantic image extrapolation method, of first generating an extrapolated segmentation map on the basis of a segmentation map from an input image, and filling the empty region for image-extension in the image with information on the basis of the extrapolated segmentation map and the input image.