Industrial Defect Image Generation With Style-Content Disentanglement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing generative adversarial networks (GANs) require large amounts of data for stable training and struggle to generate realistic images of defects on different products, especially when defective samples are scarce, limiting their effectiveness in enriching training datasets for defect detection.
Innovation Solution
A method that disentangles style and content in images using specific loss functions, allowing the generator to focus on generating novel defect patterns without altering the background, enabling the transfer of defects from one image to another, and utilizing weakly-supervised learning to enhance variance and diversity of synthetic defects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If GANs are used to generate synthetic defective images, then the training dataset can be enriched, but the generator requires large amounts of data for stable training which is not available when defective samples are scarce
Solution Approach 1:
The patent segments the image into foreground (defect) and background components, and further segments the defect into style and content components. This segmentation allows the model to learn from limited defective samples by separating the essential defect characteristics (content) from the product-specific appearance (background and style), enabling stable training even with scarce defective data.
Solution Approach 2:
The patent extracts the defect content from the background product images using a content encoder that identifies and isolates defect-related features. This extraction process separates the defect information from the product-specific background, allowing the generator to create realistic defective images without requiring large amounts of diverse defective training data.
2Adaptability or versatility
If the generator transforms the entire image style, then the target domain style is achieved, but the background is unnecessarily modified along with the defect
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different image regions: the defect region undergoes style transformation while the background region preserves its original appearance. The style encoder and content encoder work together to apply style changes locally to the defect content without modifying the background, achieving targeted style adaptation.
Solution Approach 2:
The patent introduces a content encoder as an intermediary that separates defect content from background style. This intermediary component enables the generator to transfer defect content between different product backgrounds while maintaining the appropriate style for each defect, without requiring full image-style transformation.
3Adaptability or versatility
If the generator focuses on generating diverse defect patterns, then the variance of synthetic images is improved, but the background style consistency is compromised
Solution Approach 1:
The patent segments the image processing into independent components: background processing, defect content generation, and defect style transfer. This segmentation allows the background to maintain its original style consistency while the defect region generates diverse patterns through the content encoder and style encoder working together on the defect portion only.
Solution Approach 2:
The patent applies local quality by restricting style transformations and diversity generation to the defect region only. The background maintains its original appearance and style consistency, while the defect area exhibits diverse patterns and styles, achieving both diversity and consistency simultaneously through region-specific processing.
Data Source
AI summary
A computer-implemented method of training a generator for transforming a given image according to a given target foreground domain is disclosed. A generator, discriminator, Mapping network, and a Style-Content encoder are trained. Furthermore, a method of image to image translation according to a given target foreground domain by the trained generator is provided. The generator may be trained to generate images depicting defects.