Image Domain Style Conversion for Natural Composite Foregrounds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image composition technologies often result in unnatural foreground images due to focusing solely on color elements, neglecting domain style harmonization between background and foreground images.
Innovation Solution
A machine learning model is trained to convert the domain style of a portion of an image using an encoder and decoder, allowing for the generation of a more natural composite image by maintaining the background style and adjusting the foreground style to match, utilizing blur-processed images for one-shot training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Stability of the object's composition
If image composition technology focuses only on color elements in image harmonization, then color harmony between background and foreground is improved, but the foreground image appears unnatural compared to the background image
Solution Approach 1:
The patent applies local quality by differentiating the processing of background and foreground images. The background image is processed to establish a target domain style, while the foreground image is selectively converted to match this style only in regions where naturalness is compromised. This localized approach allows color harmony to be maintained while restoring naturalness to the foreground, resolving the contradiction between color harmony and foreground naturalness.
Solution Approach 2:
The patent changes the parameter of domain style conversion by introducing a style transfer module that transforms the domain style of the foreground image to match the background image's domain style. This parameter change extends beyond color harmony to include texture, lighting, and visual characteristics, thereby improving foreground naturalness while maintaining color harmony.
2Reliability
If domain style conversion is applied to the entire image, then overall aesthetic quality is improved, but processing time and computational complexity increase
Solution Approach 1:
The patent segments the image processing into distinct stages: background image processing to establish domain style, foreground image extraction, and selective domain style conversion of the foreground. This segmentation allows the computationally intensive domain style conversion to be applied only where necessary (in the foreground) rather than the entire image, reducing processing time while maintaining aesthetic quality.
Solution Approach 2:
The patent applies partial action by converting the domain style of only the foreground image rather than the entire image. This partial conversion is sufficient to achieve the desired aesthetic quality improvement, as the background already establishes the target domain style. Applying conversion to the entire image would be excessive and unnecessarily increase processing time.
Data Source
AI summary
An image conversion method includes obtaining a first image, generating, using a machine learning model, a second image by converting a domain style of at least a portion of the first image into a first domain style, and outputting the second image. The machine learning model is trained to output a second training image of the first domain style associated with a first training image of a second domain style in response to receiving the first training image of the second domain style as input.