Pseudo Image Generation Device Annotation Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The generation of image data attached with annotation data requires a high processing load, particularly when images are obtained by imaging only a portion of a whole image or by slightly shifting the position or angle of an image, necessitating repeated attachment of annotation data.
Innovation Solution
A pseudo image generation device that transparentizes regions other than the object to be synthesized in the foreground image, synthesizes the background and foreground images, deletes overlapping background annotation data, and generates annotation data for the pseudo image, thereby reducing the need for repeated annotation attachment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If annotation data is repeatedly attached to images obtained by imaging only a portion of a whole image or by slightly shifting position or angle, then the training data coverage is improved, but the processing load increases significantly
Solution Approach 1:
The patent uses template matching to identify and copy annotation data from reference images to synthesized images. Instead of manually attaching annotation data to each variant image, the system automatically copies annotations from templates based on detected object positions, dramatically reducing the processing load while maintaining comprehensive training data coverage
Solution Approach 2:
The patent performs preliminary annotation on reference images and stores them as templates before synthesis. This preliminary action allows the system to quickly apply annotations to multiple synthesized images without repeating the full annotation process, thereby reducing overall processing time and load
2Reliability
If multiple image data in various forms is generated for AI training, then the training effectiveness is improved, but the data collection time and labor increase
Solution Approach 1:
The patent merges background images with foreground objects containing annotation data to create synthesized training images in various forms. This combining approach generates diverse training data automatically, improving training effectiveness while eliminating the need for manual collection of multiple image variants
Solution Approach 2:
The system copies annotation data from source images and applies it to synthesized images through template matching. This automatic copying mechanism generates multiple forms of training data without requiring manual data collection for each variant, significantly reducing data collection time
3Measurement precision
If annotation data is attached to each synthesized image individually, then the annotation accuracy is maintained, but the computer processing load increases
Solution Approach 1:
The patent implements template matching to copy annotation data from reference images to synthesized images. This approach maintains annotation accuracy by using proven templates while dramatically reducing computer processing load by avoiding individual annotation attachment to each synthesized image
Solution Approach 2:
The system creates universal annotation templates that can be applied to multiple synthesized images with different backgrounds and variations. This multi-functional approach maintains consistent annotation accuracy across all images while reducing the overall processing load through template reuse
Data Source
AI summary
A pseudo image generation device (10) that generates a pseudo image from foreground image data and background image data includes: a transparentizing processing unit (11) configured to transparentize a region other than an object to be synthesized in a foreground image; a synthesis unit (12) configured to synthesize the background image data and the foreground image data after transparentizing processing such that a foreground image after the transparentizing processing is superimposed on a prescribed position in a background image; a deletion unit (13) configured to delete background annotation data overlapping the object to be synthesized after synthesis from background annotation data included in the background image data and attached within the background image; and a data generation unit (14) configured to generate annotation data of the pseudo image based on the background annotation data after deletion.