Autoencoder-Guided Image Inspection Models for Imbalanced Classes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification methods using neural networks face challenges in accurately identifying abnormal products due to imbalanced training data, where normal products outnumber abnormal products, leading to biased classification and reduced accuracy in abnormality detection.
Innovation Solution
A learning model generating method that utilizes a variational autoencoder to balance data by extracting teacher data based on reconstruction errors, ensuring equal representation of normal and abnormal image classes, followed by training a convolutional neural network with this balanced data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learning is performed using imbalanced data with many normal product images and few abnormal product images, then the determination rate as normal product increases, but the accuracy of abnormal product detection decreases
Solution Approach 1:
The patent transforms the imbalanced data into balanced data by selecting samples based on reconstruction error parameters. Specifically, it selects a certain number of normal product samples with larger reconstruction errors and combines them with abnormal product samples, thereby changing the class distribution parameter from imbalanced to balanced while maintaining meaningful sample characteristics through error-based selection criteria
Solution Approach 2:
The patent applies different selection criteria to different classes of data. For normal product samples, it selects based on larger reconstruction errors (indicating harder-to-reconstruct samples that may contain subtle abnormalities), while for abnormal product samples, it selects based on smaller reconstruction errors. This local quality differentiation ensures each class contributes its most discriminative features to the balanced dataset
2Quantity of substance
If the number of normal product images is decreased to equalize class numbers, then data balance improves, but the quality of discriminative features may be lost
Solution Approach 1:
Instead of randomly decreasing the number of normal product images, the patent changes the selection parameter from random sampling to error-based sampling. It selects normal product samples with reconstruction errors above a certain threshold or among the largest errors, thereby maintaining high-quality discriminative features while achieving the desired class balance
Solution Approach 2:
The patent uses the reconstruction error as a proxy indicator to identify valuable samples. By copying the selection strategy based on error metrics rather than random selection, it ensures that the reduced set of normal product images retains the most informative characteristics for distinguishing normal from abnormal products
Data Source
AI summary
An autoencoder is caused to learn based on data of a majority class of imbalanced data. The imbalanced data is input to the autoencoder that has learned, and an error in a predetermined parameter between an input image data and an output image data is acquired. Data of each class is extracted as teacher data such that the number of pieces of first class data counted in descending order of the size of the obtained error among the first class data and the number of pieces of second class data are balanced. A convolutional neural network is caused to learn the teacher data to obtain a learning model.