Image Translation Model Using Down-Sampled Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image translation methods face a high computation burden, leading to low definition and reduced effectiveness when attempting to reduce this burden by simplifying the model or reducing image resolution.
Innovation Solution
The method involves down-sampling the original image to generate a pre-translated image, mask image, and deformation parameters, then deforming and fusing these images to produce a target translation image of the same size, utilizing deformation parameters to maintain high-definition details and reduce computation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If image translation is performed using existing methods, then translation function is achieved, but computation burden is high
Solution Approach 1:
The patent divides the image translation process into multiple stages: first translating the down-sampled image to obtain a pre-translated image, then generating deformation parameters, and finally fusing the pre-translated image with the deformed original image. This segmentation allows the heavy computation to be performed on the smaller down-sampled image while maintaining high-quality output.
Solution Approach 2:
The patent introduces a resolution dimension by processing the image at two different scales: the original high-resolution image is down-sampled to a lower resolution for translation computation, while the output is restored to the original resolution through fusion. This dimensional change reduces computation burden while maintaining quality.
2Use of energy by moving object
If image resolution is reduced to lower computation burden, then computation burden decreases, but definition and detail quality deteriorate
Solution Approach 1:
The patent performs preliminary translation on the down-sampled image to obtain a pre-translated image, then uses this pre-translated image as a guide for the final high-quality output. The deformation parameters are also pre-computed from the down-sampled image, enabling efficient final fusion without reprocessing the entire high-resolution image.
Solution Approach 2:
The down-sampled image acts as an intermediary that facilitates the translation process. It is used to compute deformation parameters and generate pre-translated images, which then guide the final fusion operation to produce the high-quality target translation image without requiring the entire high-resolution image to be processed.
3Use of energy by moving object
If model complexity is simplified to reduce computation burden, then computation burden decreases, but translation effectiveness reduces
Solution Approach 1:
The patent applies different processing qualities to different parts of the pipeline: the down-sampled image is processed with standard translation, while the final fusion operation uses high-quality deformation parameters and mask images to ensure the critical regions (such as facial features) maintain high translation effectiveness despite the reduced computation on the down-sampled version.
Data Source
AI summary
A method for translating an image, a method for training an image translation model, and related electronic devices are proposed. In the method for translating an image, an image translation request carrying an original image is obtained. A down-sampled image is generated by down sampling the original image. A pre-translated image, a mask image, and deformation parameters are generated based on the down-sampled image. A size of the pre-translated image and a size of the mask image are the same as a size of the original image. A deformed image is obtained by deforming original image based on the deformation parameters. The deformed image, the pre-translated image and the mask image are fused to generate a target translation image.