Simulated Environment Synthetic Image Generation for ML Training

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current methods for training machine-learning models to identify real-world elements from images or video sequences are labor-intensive, slow, and fail to generalize effectively, as they either require extensive manual tagging of real-world data or include irrelevant information when using synthetic data.

Innovation Solution

A method using a simulated environment to generate realistic and abstract elements based on appearance parameters, creating synthetic images that are then used to train machine-learning models to identify real-world elements, allowing for quick, generalized, and labor-efficient training without human intervention.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If real-world video footage and discrete images are used for training, then diversity of training data is improved, but labor intensity and training time increase significantly

Engineering Contradiction:

Improvediversity of training dataVSAvoidtraining speed

Core Design Contradiction:

Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent uses synthetic image generation to create copies of real-world scenes with controlled variations. Instead of manually collecting and tagging diverse real-world images, the system generates synthetic images that replicate real-world conditions while automatically providing ground truth labels, thus improving training diversity without increasing labor intensity.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The system performs preliminary scene understanding and element detection on synthetic images before they are used for training. By pre-processing and pre-tagging synthetic images with accurate labels during generation, the system eliminates the need for manual tagging of diverse real-world images, thereby improving both diversity and training efficiency.

Inventive Principle:

Principle #10Preliminary action

2Productivity

If synthetic images are used for training, then training speed is improved, but generalization capability deteriorates

Engineering Contradiction:

Improvetraining speedVSAvoidgeneralization capability

Core Design Contradiction:

ProductivityVSAdaptability or versatility

Solution Approach 1:

The patent systematically varies multiple parameters in synthetic image generation including lighting conditions, camera angles, object positions, backgrounds, and weather conditions. By changing these parameters to create diverse synthetic training data, the model learns to generalize across different real-world conditions while maintaining fast training speeds.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The synthetic image generation system serves multiple functions: it creates diverse training data, provides automatic ground truth labels, controls for irrelevant information, and enables systematic parameter variation. This multi-functionality allows the system to improve generalization capability while maintaining training speed advantages.

Inventive Principle:

Principle #6Universality (Multi-functionality)

3Productivity

If synthetic images with specific element designs are used, then training efficiency is improved, but class specificity increases causing exclusion of real-world variants

Engineering Contradiction:

Improvetraining efficiencyVSAvoidclass generalization

Core Design Contradiction:

ProductivityVSAdaptability or versatility

Solution Approach 1:

The system changes design parameters of elements in synthetic images including colors, shapes, sizes, patterns, and configurations. By generating synthetic images with varied element designs rather than fixed designs, the model learns to recognize the essential characteristics of element classes while remaining adaptable to real-world variants during deployment.

Inventive Principle:

Principle #35Parameter changes

4Productivity

If synthetic images with repetitive backgrounds are used, then generation speed is improved, but relevance accuracy deteriorates due to inclusion of irrelevant information

Engineering Contradiction:

Improveimage generation speedVSAvoidrelevance accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent varies background parameters in synthetic image generation including background types, textures, colors, and configurations. By changing background parameters to create diverse and realistic backgrounds rather than repetitive ones, the system maintains generation speed while improving relevance accuracy by preventing the model from associating irrelevant repetitive patterns with target elements.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS10242264B1System and method for training a machine-learning model to identify real-world elements

Publication Date: 2019.03.26 IMAGRY ISRAEL LTD

AI summary

A method and a system for training a machine-learning model to identify real-world elements using a simulated environment (SE) may include (a) receiving at least one set of appearance parameters, corresponding to appearance of real-world element; (b) generating one or more realistic elements, each corresponding to a variant of at least one real-world element; (c) generating one or more abstract-elements; (d) placing the elements within the SE; (e) producing at least one synthetic image from the SE; (f) providing the at least one synthetic image to a machine-learning model; and (g) training the machine-learning model to identify at least one real-world element from the at least one synthetic image, that corresponds to at least one realistic element in the SE.