cGAN Training with Embedded Parameter Encodings for Sparse Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conditional Generative Adversarial Networks (cGANs) face challenges in training accuracy when labelled training data is sparse, hindering their performance in applications where data augmentation is necessary.
Innovation Solution
A method for training cGANs by embedding graphical encodings of physical parameter values into images, using a loss function that measures the difference in these encodings, and incorporating calibration graphics to improve training accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If graphical encodings are embedded into images to enhance training accuracy, then measurement precision of physical parameters is improved, but device complexity increases
Solution Approach 1:
The patent merges graphical encodings directly into the image data itself, combining the visual information with the parameter encoding information in a single integrated structure. This allows the neural network to process both the image content and the parameter values simultaneously without requiring separate processing streams, thereby improving training accuracy while controlling the increase in processing complexity through unified data representation.
2Measurement precision
If multiple graphical encodings are embedded per image to capture multiple parameters, then measurement precision is improved, but data processing time increases
Solution Approach 1:
The patent represents multiple physical parameters by embedding multiple graphical encodings into different spatial locations or layers of the image data structure. This dimensional expansion allows the neural network to simultaneously encode and process multiple parameter values within the same training iteration, improving measurement precision while managing training time through efficient multi-dimensional data representation rather than sequential processing.
3Measurement precision
If calibration graphics are added to images to improve encoding accuracy, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent introduces calibration graphics as intermediary elements within the image data that serve as reference markers for encoding calibration. These calibration graphics act as mediators between the physical parameters and the neural network's encoding system, enabling accurate calibration of graphical encodings without requiring complex external calibration procedures or separate processing systems. The calibration graphics are integrated directly into the image structure, simplifying the overall system architecture.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method of training a conditional Generative Adversarial Network (cGAN) is disclosed. The cGAN (102) has a generator (305) and a discriminator (306). The method comprises obtaining a collection of images of components (301) , each image having a physical parameter value (302) relating to the component associated therewith, for each one of the collection of images of components, embedding a plurality of graphical encodings into the image that encode the associated physical parameter value, and training the cGAN using the collection of images of components with their embedded plurality of graphical encodings.