Image Domain Style Conversion for Natural Composite Scenes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image composition technologies focus solely on color elements, leading to unnatural foreground images compared to the background, necessitating the development of image conversion technology that considers domain style for more natural composite images.
Innovation Solution
An image conversion method using a machine learning model that converts the domain style of a portion of an image into a desired style, employing an encoder to generate feature vectors and a decoder to output images with the desired domain style, trained using blurred images for one-shot learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional image composition technology is used to process only color elements, then the processing is simple and fast, but the foreground image looks unnatural compared to the background image
Solution Approach 1:
The patent changes the processing parameters from only color elements to include domain style parameters (blur level, depth of field, lens characteristics). The machine learning model processes images based on these style parameters to generate natural composite images that match the domain characteristics of the background image.
Solution Approach 2:
The patent introduces a machine learning model as an intermediary between the foreground image and background image. This model learns the domain style relationships from training data and mediates the composition process to ensure natural transitions between foreground and background while maintaining their respective domain characteristics.
2Manufacturing precision
If image composition focuses only on color elements, then the processing time is short, but the composite image lacks naturalness
Solution Approach 1:
The patent performs preliminary action by pre-training the machine learning model on a large dataset of images with known domain styles. This pre-training allows the model to capture complex style relationships in advance, so that during actual image composition, the model can quickly generate natural results without requiring real-time complex processing.
Solution Approach 2:
The patent changes the approach from processing only color elements to processing domain style parameters. By focusing on style transfer rather than detailed pixel-level processing, the system achieves natural composite images with acceptable processing times through efficient machine learning inference.
3Adaptability or versatility
If conventional image composition is used, then the processing is simple, but the foreground image does not match the domain style of the background image
Solution Approach 1:
The patent introduces domain style parameters as new processing dimensions. Instead of only adjusting color elements, the system now processes blur levels, depth of field, lens characteristics, and other domain-specific parameters to match the background image's domain style with the foreground image.
Solution Approach 2:
The machine learning model serves as an intermediary that learns domain style relationships from training data. It mediates between the foreground and background images by applying appropriate style transformations, enabling the system to adapt to different domain styles without requiring complex manual processing rules.
Data Source
Figure 1
Figure 2
Figure 3
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.