Synthetic Image Generation for Object Detection Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Developing large realistic visual data sets with high accuracy, specificity, and diversity for object detection systems is challenging, especially using conventional methods that rely on manual data collection and human labeling, which slows the development and deployment of these systems.
Innovation Solution
An unsupervised cross-domain synthetic image generation system that uses a processor and memory to generate augmented CAD models based on 3D CAD models, incorporating a graph data structure and generative adversarial networks to produce synthetic photorealistic images with attributes corresponding to various classes, including anomalies, thereby creating diverse and accurate training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional manual data collection and human labeling methods are used, then data accuracy can be achieved, but productivity is significantly reduced
Solution Approach 1:
The patent uses generative adversarial networks to create synthetic copies of real-world images and data patterns. The GAN generates photorealistic training images that replicate the statistical properties and visual characteristics of actual data without requiring manual collection or labeling, thus achieving both accuracy and high productivity
Solution Approach 2:
The system employs unsupervised learning where the model automatically learns data distributions and generates training samples without human intervention. The unsupervised anomaly detection framework self-trains on the generated data, eliminating the need for manual labeling while maintaining data quality and accuracy
2Adaptability or versatility
If diverse training data with various view angles and environmental conditions is collected manually, then data diversity is improved, but the complexity of the data collection process increases
Solution Approach 1:
The patent systematically varies multiple parameters including view angles, lighting conditions, environmental settings, and object orientations during synthetic data generation. By programmatically adjusting these parameters, the system achieves comprehensive data diversity without the logistical complexity of manual data collection across different physical scenarios
Solution Approach 2:
The generative adversarial network serves multiple functions simultaneously: it generates diverse images, applies various transformations, creates different environmental conditions, and produces annotated training data all through a single unified system, eliminating the need for separate data collection processes for each type of variation
3Measurement precision
If large scale realistic visual data sets are created manually, then measurement precision is improved, but loss of time increases
Solution Approach 1:
The system performs preliminary synthetic data generation and model pre-training before actual deployment. By pre-generating large volumes of diverse training data and pre-training the detection model on synthetic datasets, the system reduces the time required for final deployment and achieves detection accuracy without the time-consuming manual data collection process
Solution Approach 2:
The generative adversarial network rapidly produces large volumes of photorealistic training images that copy the essential characteristics of real-world data. This synthetic data copying process occurs computationally at high speed, generating millions of training samples in minutes rather than the months or years required for manual data collection and labeling
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image generation system (10) is provided to: receive a 3D CAD (computer aided design) model (28) comprising 3D model images of a target object; generate a graph data structure (23); based on the 3D CAD model (28) of the target object and the graph data structure (23), generate a plurality of augmented CAD models (32a-d) of the target object comprising a plurality of data sets (32a-d), each data set (32a-d) respectively corresponding to an associated one of a plurality of attribute classes, each data set (32a-d) comprising a plurality of 2D model images; input the plurality of data sets (32a-d) into a generative adversarial network (34); generate synthetic photorealistic images (40) of the target object using the plurality of generators (36a-d) of the generative adversarial network (34), the synthetic photorealistic images (40) including attributes in accordance with the plurality of data sets (32a-d) corresponding to the plurality of attribute classes; and output the synthetic photorealistic images (40) of the target object.