Synthetic Training Image Generation for Object Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object recognition systems face challenges in adapting to new environments due to the need for extensive manual labeling and recording of training images under varying imaging conditions, which is time-consuming and resource-intensive.

Innovation Solution

A method for generating training image data by modifying generic image data with respect to imaging-related parameters, such as fisheye distortion and noise, and determining similarity using Kernel Principal Component Analysis (KPCA) and Local Binary Patterns (LBP), allowing for the automatic creation of training images that mimic real-life conditions without manual labeling.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If recorded image data is collected under various imaging conditions and manually labelled, then the object recognition system can be trained to recognize objects under different conditions, but the process becomes time-consuming and resource-intensive

Engineering Contradiction:
Improveobject recognition reliabilityVSAvoidtraining data preparation time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent uses a template image as a base and generates multiple modified copies by applying different imaging conditions (distortion, noise, brightness, contrast) to create synthetic training images. This copying approach eliminates the need to manually collect and label numerous real images under various conditions, significantly reducing time and resource requirements while maintaining training effectiveness

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system pre-defines a template image representing the object type and pre-establishes various imaging condition parameters (fisheye distortion, noise levels, brightness variations). These preliminary preparations allow for rapid generation of diverse training images without needing to physically capture images under each condition, addressing the time-consuming nature of traditional data collection

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If a larger number of training images are provided to improve machine learning, then the system can better handle varying imaging conditions, but the manual selection and labelling process becomes more resource-intensive

Engineering Contradiction:
Improveadaptation to imaging conditionsVSAvoiddata processing complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system automatically generates training images by applying predefined imaging transformations to a template image without requiring manual selection or labelling. The automated process handles all variations of imaging conditions (distortion, noise, lighting) through algorithmic transformations, eliminating the need for human intervention in data preparation and significantly reducing operational complexity

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent systematically varies imaging parameters (fisheye distortion strength, noise intensity, brightness levels, contrast) to generate diverse training images from a single template. This parameter-based approach allows for controlled generation of numerous training variations without manually processing each image, reducing both time and computational resources required

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If additional sets of training images are collected for new object types, then the system can recognize new object types, but the manual labelling process must be repeated for each new type

Engineering Contradiction:
Improvenew object type recognitionVSAvoidtraining efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The system uses a universal template image approach where a single template can generate training images for multiple object types by applying different transformations and configurations. This universal method allows the same generation process to serve multiple object types, eliminating the need to repeat manual data collection and labelling for each new object type and significantly improving training efficiency

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3343432B1Generating training images for machine learning-based object recognition systems
Publication Date: 2024.03.20 ELEKTROBIT AUTOMOTIVE GMBH
  • EP3343432B1 patent drawingFigure 1~2
  • EP3343432B1 patent drawingFigure 3
  • EP3343432B1 patent drawingFigure 4~5

AI summary

A method, performed by a computing device, for generating training image data for a machine learning-based object recognition system is described. The method comprises receiving generic image data of an object type, receiving recorded image data related to the object type, and modifying the generic image data with respect to at least one imaging-related parameter. The method further comprises determining a degree of similarity between the modified generic image data and the recorded image data, and, when the determined degree of similarity fulfills a similarity condition, storing the modified generic image data as generated training image data of the object type. Further described are a computing device, a computer program product, a system and a motor vehicle.