Self-Supervised Image Outpainting for Defect Data Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in manufacturing electronic devices, such as OLEDs, is the inefficiency and high cost of human-operated defect identification and classification due to a lack of balanced defect-free and defective image data for AI training, particularly in newer products where defect images are scarce.
Innovation Solution
An AI-based outpainting model using self-supervised learning generates synthetic defect images by expanding and cropping existing images to create a larger dataset, including both defect-free and defective images, thereby enhancing training data for defect classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If human operators manually identify and classify defects, then defect detection can be performed, but the process becomes inefficient and costly due to lack of balanced training data for AI systems
Solution Approach 1:
The patent uses generative AI models to create synthetic copies of defect images. The system generates realistic defect images by copying and transforming existing defect patterns, creating balanced training datasets without requiring additional physical defect samples. This allows AI models to be trained efficiently on synthesized data that mimics real defect variations.
Solution Approach 2:
The system transforms existing defect images by changing various parameters including rotation angles, scaling factors, brightness levels, contrast values, and noise levels. These parameter transformations generate diverse training samples from limited original defect images, creating balanced datasets that improve AI model training efficiency while reducing the need for manual defect collection.
2Measurement precision
If AI models are trained with limited defect images, then training can proceed, but the model accuracy and generalization capability are insufficient
Solution Approach 1:
The patent segments the training data generation process into multiple independent transformation operations. Each defect image undergoes separate transformations including geometric transformations (rotation, scaling), photometric transformations (brightness, contrast adjustments), and noise additions. This segmentation allows systematic generation of diverse training samples while maintaining control over each transformation parameter.
Solution Approach 2:
The system adds new dimensions to the training data by creating multi-scale versions of defect images at different resolutions, generating images with various noise levels and conditions, and creating augmented datasets that extend beyond the original single-scale, single-condition images. This dimensional expansion significantly increases the effective training data volume.
3Quantity of substance
If more defect images are collected from production, then training data volume increases, but the process remains inefficient and costly
Solution Approach 1:
The system implements self-service by enabling AI models to automatically generate their own training data. Instead of relying on external sources or manual collection processes, the generative AI model uses existing defect images and transformation algorithms to autonomously create balanced training datasets, eliminating the need for continuous manual defect collection and annotation processes.
Solution Approach 2:
The patent applies preliminary transformations to defect images during the data preparation phase, creating augmented training datasets before model training begins. By pre-generating diverse defect variations through geometric and photometric transformations, the system prepares comprehensive training data in advance, avoiding the need for ongoing data collection during model development and deployment.
Data Source
AI summary
A method may include receiving, by a processor comprising an outpainting model, an input NG image, generating a first set of masks based on the input NG image, and applying the first set of masks to the input NG image to expand a size of the input NG image by outpainting a first region relative to the input NG image to generate an output NG image.