Two-Step Diffusion for High-Resolution Synthetic Display Defect Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current defect detection in display devices like OLEDs and QD-OLEDs is time-consuming, expensive, and prone to human error due to the imbalance in the number of defect-free and defective samples, necessitating a more efficient and robust artificial intelligence-based classification and repair system.
Innovation Solution
A two-step method involving training multi-class-conditioned diffusion models to generate high-resolution synthetic defect images by cropping and superimposing defect images, utilizing connected components algorithms to identify regions of interest, and training AI classifiers for defect detection and auto-repair.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a robust AI defect detection system is implemented, then defect detection accuracy is improved, but the computational power requirements and training complexity increase
Solution Approach 1:
The patent segments the image generation process into two distinct stages: a first training stage that generates intermediate defect images, and a second training stage that generates final high-resolution synthetic defect images. This segmentation allows the system to achieve high detection accuracy while managing computational complexity by dividing the training workload into manageable phases with different resource requirements.
2Measurement precision
If high-resolution synthetic defect images are generated directly, then defect detection precision is improved, but the training time and computational resources increase significantly
Solution Approach 1:
The patent performs preliminary actions by first generating intermediate defect images and establishing a baseline defect detection model before proceeding to generate high-resolution synthetic defect images. This preliminary training phase prepares the system by learning general defect patterns, which accelerates the subsequent high-resolution image generation and reduces overall training time while maintaining defect detection precision.
3Reliability
If more defect images are used for training, then defect classification robustness is improved, but the data availability decreases since defective samples are only 1-2% of total production
Solution Approach 1:
The patent creates synthetic copies of defect images through a two-stage diffusion model training process. The first stage generates intermediate defect images that capture essential defect characteristics, and the second stage produces high-resolution synthetic defect images that preserve these characteristics while providing abundant training data. This copying approach enables robust defect classification by generating sufficient defect samples without requiring additional physical defective products.
Solution Approach 2:
The patent changes the resolution parameter between training stages by generating intermediate images at lower resolution in the first stage and then generating high-resolution synthetic defect images in the second stage. This parameter transformation allows the system to learn defect patterns efficiently at lower computational cost initially, then enhance image quality for final robust classification, effectively overcoming the limitation of scarce defect samples.
Data Source
AI summary
A method of generating high-resolution synthetic images includes training a first diffusion model on a corpus of real images of a first class and real images of a second class; generating a synthetic image of the first class from a real image of the second class, and a synthetic image of the second class from a real image of the first class; determining at least one region of interest of the real images of the second class; cropping the real images of the first class and the real images of the second class based on the region of interest to generate a corpus of cropped images; training a second diffusion model on the cropped images; generating synthetic images of the first class from the corpus of cropped images;and superimposing the images of the first class on the real images of the second class to generate the high-resolution synthetic images.