Vision Inspection Model Generation Using Similar Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based vision inspection systems for product production lines face challenges in generating accurate new learning models with limited training data, as the determination accuracy is low due to small amounts of defective data, and require separate models for each process and product type, leading to inefficiencies and increased time in determining pre-training models.
Innovation Solution
An AI-based new learning model generation system that extracts candidate data sets based on determination type information, calculates similarity between training images using feature maps and feature point distribution, and selects an optimal pre-training model to increase determination accuracy and efficiency by utilizing existing training data sets, thereby generating a new learning model with improved accuracy even with small amounts of training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If a new learning model is generated using only new training data with small amounts of defective data, then the model can be quickly created for new products, but the determination accuracy is low
Solution Approach 1:
The system performs preliminary action by selecting and preparing relevant training data from historical data sets before generating the new learning model. The data selection module pre-identifies historical data that matches the new product's characteristics, and the training data generation module pre-processes this data into appropriate formats, so that when new training data is limited, the model can still achieve high accuracy without extensive data collection time
Solution Approach 2:
The system merges new training data with selected historical training data to create a combined training data set. The data selection module identifies historical data sets with similar product characteristics, and the training data generation module combines these with new data, ensuring the merged data provides sufficient defective examples for accurate model training while maintaining quick generation
2Measurement precision
If separate learning models are generated for each process and product type, then the determination accuracy for specific defects is improved, but the device complexity and time for model determination increase
Solution Approach 1:
The system implements universality by creating a single learning model generation system that can handle multiple product types and processes. The data selection module automatically identifies relevant historical data based on product and process characteristics, allowing one versatile model to perform specialized inspection tasks across different products without requiring separate dedicated models for each
3Measurement precision
If separate learning models are generated for each process and product type, then the inspection accuracy is improved, but the time required for determining pre-training models increases
Solution Approach 1:
The system performs preliminary action by pre-selecting and organizing historical training data according to process and product type classifications. The data selection module pre-identifies relevant historical data sets based on similarity metrics, so when a new product or process needs inspection, the appropriate pre-training data is already prepared and ready, significantly reducing the time to determine pre-training models while maintaining high inspection accuracy
Data Source
AI summary
An AI-based new learning model generation system for vision inspection on a product production line is proposed. In the AI-based new learning model generation system, the candidate set extraction module extracts two or more candidate data sets on the basis of determination type information from among a plurality of training data sets that have been applied to learning of existing learning models previously generated for the vision inspection on the product production line. In addition, an additional set determination module calculates similarity between training images of new training data and a candidate data set, and determines any one greater than or equal to a reference value as an additional training data. In addition, the new model generation module may generate a new learning model by training the additional training data set and the new training data as a pre-training model.