Diffusion Model Sampling for Synthetic Defect Image Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in manufacturing processes is the scarcity of defect images for new products, which hinders the development of robust defect detection classifiers, especially when defect samples constitute only a small subset of the total production.
Innovation Solution
Utilizing AI-based generative models, specifically diffusion model sampling with automatic masking, to create synthetic defect images for target products by learning from defect-free and defective images of source products, thereby overcoming the data imbalance issue.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If AI-based generative models are used to create synthetic defect images, then the quantity of defect images is improved, but the quality and accuracy of the generated images deteriorates
Solution Approach 1:
A mask is introduced as an intermediary element during the diffusion model sampling process. The mask selectively controls which regions of the image receive defect characteristics, allowing the background to remain high-quality while defects are generated in specific targeted areas. This resolves the contradiction by mediating between the need for synthetic defect generation and the need for background quality preservation.
Solution Approach 2:
The patent applies different qualities to different regions of the image: the background maintains high quality from the original defect-free image, while specific local regions receive defect characteristics through the masking process. This local differentiation allows the system to generate synthetic defect images with both quantity improvement and localized quality control.
2Measurement precision
If manual masking is used to generate synthetic defect images, then the location accuracy of defects is improved, but the reliability of defect generation deteriorates
Solution Approach 1:
The system performs self-service by automatically generating the mask through the diffusion model sampling process itself, rather than requiring external manual creation. The mask is derived from the model's own predictions during sampling, making the process self-contained and more reliable while maintaining location accuracy through the model's learned defect patterns.
Solution Approach 2:
The mask generation is performed as a preliminary action during the early diffusion sampling steps, before the final defect image is generated. This preliminary mask creation allows the system to pre-determine defect locations based on learned patterns, ensuring both location accuracy and generation reliability in subsequent sampling steps.
3Productivity
If diffusion model sampling is performed without masking, then the productivity of image generation is improved, but the manufacturing precision of the background deteriorates
Solution Approach 1:
The mask serves as an intermediary that separates the defect generation process from the background processing. By introducing this selective mechanism, the system maintains high productivity through automated diffusion sampling while preserving background quality by preventing unwanted modifications to non-defect regions.
Solution Approach 2:
The masking approach applies local quality control by allowing unrestricted diffusion sampling in defect regions while protecting background regions from degradation. This spatial differentiation enables the system to maintain both high productivity through full automation and high background quality through selective protection.
Data Source
AI summary
A system and a method are disclosed for defect image generation using diffusion model sampling. The method includes generating, by a processor via a diffusion model, a noisy image from a defect-free image, generating, by the processor via the diffusion model, a sampled defect image and a sampled defect-free image from the noisy image, generating, by the processor, a mask based on the sampled defect image and the sampled defect-free image, generating, by the processor, a synthetic defect image by generating an additional sampled defect image based on the noisy image and the mask, and transmitting, by the processor, the synthetic defect image.