Teacher Data Collection Using VAE Latent Coverage for Defect Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing inspection devices face challenges in efficiently collecting high-quality teacher data for defect classification due to the scarcity of defective products and the need for expert input, leading to inefficiencies in time and cost, and inaccuracies with non-expert data.
Innovation Solution
A method using a variational autoencoder to encode expert and non-expert data into one-dimensional latent variables, comparing the differences in these variables to determine when sufficient non-expert data collection is complete, minimizing the need for expert data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If expert data is used to collect teacher data, then data quality and reliability are improved, but collection time and cost increase significantly
Solution Approach 1:
The patent introduces a variational autoencoder as an intermediary system that learns from expert data and then generates synthetic defective product images. This intermediary model transfers the knowledge from limited expert data to generate additional training data, reducing the need for extensive manual expert labeling while maintaining data quality
Solution Approach 2:
The patent performs preliminary training of the variational autoencoder using expert data before actual teacher data collection. This preliminary action establishes a trained model that can then generate synthetic data, allowing the system to be prepared in advance and reducing the time needed during actual data collection phases
2Productivity
If non-expert data is used to collect teacher data, then collection efficiency is improved, but data quality and reliability deteriorate
Solution Approach 1:
The trained variational autoencoder acts as an intermediary that processes and generates data based on expert-level patterns. When non-experts provide input data, the model processes it through learned expert patterns to generate high-quality synthetic defective product images, ensuring data reliability even when collected by non-experts
Solution Approach 2:
The patent uses the variational autoencoder to copy and replicate the characteristics of expert-labeled data. By learning the underlying patterns from expert data, the model can generate synthetic copies that maintain the quality and reliability features of expert data while being produced more efficiently
3Measurement precision
If more defective product data is collected, then classification model accuracy is improved, but the scarcity of defective products makes collection difficult
Solution Approach 1:
The patent uses the variational autoencoder to generate synthetic copies of defective product images. By copying and transforming the features from limited real defective products, the system creates additional training data that increases the effective quantity of available data without requiring more physical defective products
Solution Approach 2:
The patent applies parameter changes by transforming the latent space representations generated by the variational autoencoder. By manipulating the latent variables and decoding them, the system generates diverse synthetic images with varied parameters while maintaining the essential characteristics of defective products, thereby increasing data quantity and diversity
Data Source
AI summary
A teacher data collecting method in a defect classification model for classifying a defect by using, as teacher data, a few pieces of expert data and many pieces of non-expert data, includes: encoding, into one dimension, a latent variable of a variational auto encoder that has been caused to perform learning the expert data; inputting the non-expert data into the variational auto encoder and encoding a latent variable into one dimension; calculating maximum values and minimum values of the latent variable in one dimension of the expert data and the non-expert data; and determining whether to complete collection of the non-expert data, based on a ratio of a difference between the maximum value and the minimum value of the latent variable in one dimension of the non-expert data to a difference between the maximum value and the minimum value of the latent variable in one dimension of the expert data.