Defect Inspection System Using Multiple Non-Defective Image Estimation Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing defect inspection systems using deep learning models, such as those employing Variational Auto Encoders (VAE), face challenges in reproducibility due to random processing during model learning, leading to varying inspection results even with the same image dataset.
Innovation Solution
A defect inspection system and method that utilize a multiple non-defective product image estimation model trained with a learning unit, which captures images of samples, estimates multiple non-defective product images for one input image, and extracts defective parts using these estimated images, thereby enhancing reproducibility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a deep learning model with random processing is used for defect inspection, then the model can learn from image datasets, but the inspection results vary depending on the learning result of each model
Solution Approach 1:
The patent divides the single model into multiple models (first learning model and second learning model) that perform different functions: one reconstructs images while the other generates defect candidate images. This segmentation allows each model to specialize, reducing the variability and improving the consistency and reliability of inspection results through model specialization.
Solution Approach 2:
The patent combines the outputs of multiple learning models into a unified inspection process. The defect detection unit integrates information from both the reconstructed image and the defect candidate image to make final defect detection decisions, merging multiple perspectives to improve overall inspection consistency and reliability.
2Ease of manufacture
If unsupervised learning is used to generate reconstructed images, then the model can process image data without labels, but random processing during learning causes varying internal parameters
Solution Approach 1:
The patent segments the unsupervised learning process into two distinct models with different objectives: one for image reconstruction and another for defect candidate generation. This allows each model to learn specific patterns without the randomness affecting both models equally, improving output consistency while maintaining training ease.
Solution Approach 2:
The patent changes the learning parameters and objectives for different models to reduce randomness effects. By assigning different loss functions and training criteria to each model, the system achieves more consistent and reliable outputs while maintaining the benefits of unsupervised learning.
3Device complexity
If a single learning model is used for defect inspection, then the system is simple, but inspection results lack reproducibility due to random learning processes
Solution Approach 1:
The patent segments the inspection function into multiple specialized models rather than using a single general-purpose model. This segmentation improves reliability and reproducibility by reducing the impact of random initialization and learning variations, while the modular architecture keeps the added complexity manageable.
Data Source
AI summary
Provided is a defect inspection system 1 having an imaging device 100 for acquiring an observation image of a sample 107, including a learning unit 330a that trains a multiple non-defective product image estimation model 401 that captures an image of the sample 107 to acquire a learning image and estimates a plurality of non-defective product images of the sample 107 for one input image using the learning image, and a defect inspection unit 330b that captures an image of an inspection target sample using the imaging device 100 to acquire an inspection target image 1000, inputs the inspection target image 1000 into the multiple non-defective product image estimation model 401 trained, outputs a plurality of estimated non-defective product images 440 corresponding to the inspection target image 1000, and extracts a defective part using the plurality of estimated non-defective product images 440.