AI Visual Analysis Training Data Generation via Depth-Aware Image Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence visual analysis models require extensive training data, which is difficult to obtain, especially in inaccessible environments like military facilities or hazardous areas, and complex networks are often not feasible for real-time object detection.
Innovation Solution
A training data generation method and apparatus that involves receiving a background image and object-related information, generating a target object image, estimating depths using a depth estimation model, determining insertion locations and scales, synthesizing images using a multi-scale visual analysis model, and generating final training images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If complex networks are used to improve detection accuracy and handle various object types, then the model can detect more diverse objects, but the amount of training data required increases significantly and processing speed decreases
Solution Approach 1:
The patent uses image synthesis technology to generate synthetic training images by copying and transforming real images through virtual scene construction. This creates artificial training data without requiring additional real-world data collection, directly addressing the training data scarcity problem while maintaining model versatility
Solution Approach 2:
The system pre-processes real images to extract objects and backgrounds, then pre-synthesizes multiple variations of training images before actual model training. This preliminary data preparation reduces the need for extensive real-time data collection and accelerates the training process
2Adaptability or versatility
If complex networks are used to improve detection accuracy, then the model can handle diverse objects, but processing speed decreases making real-time detection difficult
Solution Approach 1:
The patent applies data augmentation techniques that generate more training samples than strictly necessary (excessive action), allowing the model to achieve good performance with simpler network architectures. This partial approach to comprehensive data coverage enables faster processing while maintaining detection capability
Solution Approach 2:
The training data synthesis process is segmented into independent steps: object extraction, background selection, virtual scene construction, and image composition. This modular approach allows efficient processing and enables the use of lighter network models by providing sufficiently diverse training data without requiring complex architectures
3Adaptability or versatility
If extensive training data is collected from real environments, then the model can learn diverse object scenarios, but data collection becomes difficult or impossible in inaccessible environments
Solution Approach 1:
The system copies existing real images and synthesizes virtual scenes by combining extracted objects with synthetic or real backgrounds. This creates unlimited training data variations without requiring physical access to diverse real-world environments, solving the data collection accessibility problem
Solution Approach 2:
The patent introduces an image synthesis system as an intermediary between limited real data and the need for extensive diverse training data. This intermediary process generates intermediate synthetic images that bridge the gap, enabling model training for inaccessible environments without direct data collection there
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Disclosed are a training data generation method and apparatus for an artificial intelligence visual analysis model. The training data generation method includes: (a) receiving a background image and object-related information; (b) generating a target object image that meets a condition based on the object-related information; (c) applying the background image to a depth estimation model to estimate depths of each pixel location and generate a depth map; (d) determining an insertion location and scale at which the target object image is inserted based on the depth map to generate a local object insertion image and a mask image, respectively; (e) simply synthesizing the local object insertion image and the background image and then generating a synthesis image reflecting a scale feature of the background image through a trained multi-scale visual analysis model; and (f) generating a final training image using the background image, the synthesis image, and the mask image.