Visual Inspection Image Classification Using Style-Transferred Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for image classification face challenges when trained with data having divergent statistical properties, leading to decreased performance and misclassification, particularly in binary classification schemes, and the use of synthetic data is limited due to statistical discrepancies with real images.
Innovation Solution
A method involving a generative machine learning model to transfer image style between training images, preserving content while altering style, enabling the use of computer-generated images for training, and a deep neural network for classification, optimizing the model for real-world image styles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If training data with widely divergent statistical properties is used, then the quantity and diversity of training data is increased, but the model's ability to generalize decreases and performance drops sharply
Solution Approach 1:
The patent applies parameter changes by transforming the statistical properties of training images to match the target domain. A style transfer network modifies parameters such as color distribution, texture patterns, and lighting characteristics of source domain images to align with the target domain's statistical properties, enabling the model to generalize across domains with divergent characteristics
Solution Approach 2:
The patent introduces a style transfer network as an intermediary component between the source domain training data and the target domain application. This intermediary transforms the statistical properties of training images to bridge the gap between domains, allowing the classification model to learn from diverse data while maintaining generalization ability
2Loss of time
If synthetic image data is generated for training, then the cost and time of data acquisition is reduced, but the statistical properties of synthetic images differ greatly from real images
Solution Approach 1:
The patent uses copying by generating synthetic images that replicate the essential structures and features of real images. These synthetic copies are then processed through a style transfer network that copies the statistical properties (color distributions, texture patterns, lighting characteristics) from real target domain images, creating training data that both reduces acquisition costs and maintains statistical fidelity
Solution Approach 2:
The patent applies parameter changes by systematically modifying the statistical parameters of synthetic images to match those of real images. The style transfer network adjusts parameters including color histograms, texture frequency spectra, and illumination patterns, transforming synthetic images to possess the same statistical characteristics as real target domain images
3Stability of the object's composition
If training data is created under controlled conditions with low variance, then the consistency and quality of training data is improved, but the model fails to generalize to images with different statistical properties
Solution Approach 1:
The patent applies dynamics by making the training data statistically adaptable rather than static. The style transfer network dynamically adjusts the statistical properties of controlled training images to match various target domain characteristics, enabling the model to learn from consistent source data while adapting to diverse target conditions during inference
Solution Approach 2:
The patent uses parameter changes to transform the fixed statistical properties of controlled training data into variable properties that can match different target domains. By modifying parameters such as color distributions, texture patterns, and lighting conditions through style transfer, the model learns from consistent source data while becoming adaptable to varying target conditions
Data Source
Figure 1
Figure 2
Figure 3(a)~3(c)
AI summary
A computer-implemented method of generating training data for a first machine learning model for image classification uses an image style transformation to transform a set of first images having a first image style into a corresponding set of images with preserved image content but of a second image style. The image style transformation may particularly be performed using a GAN. The transformed images are used as training data to train a first machine learning model for image classification according to a defined classification scheme. The trained first machine learning model is then used for inspecting images, e.g., of products during or after their manufacturing, and for classifying them according to the classification scheme. Further provided are a data processing system comprising a computer program implementing the training methods is used for the training, and an inspection system comprising a computer program implementing the inspection method.