AI Vision Inspection Synthetic Defect Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The scarcity of defective product images in production sites hinders the development of effective vision inspection systems, which rely on balanced datasets to accurately distinguish between normal and defective products, leading to variability in accuracy and efficiency due to operator-dependent vision inspection methods.
Innovation Solution
An artificial intelligence device that uses an image restoration model and feature extraction model to generate additional training data by modifying normal product images, allowing for the distinction between normal and defective products based on image analysis, and determines product defects by calculating the distance between expression vectors of inspection images and restored images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If vision inspection systems are developed using actual defective product images, then measurement precision improves, but the quantity of available training data deteriorates due to low defect frequency in production
Solution Approach 1:
The patent generates synthetic defective product images by copying and modifying normal product images through various transformation operations (adding defects, changing colors, rotating, cropping) to create artificial training data that mimics real defective products without requiring actual defective samples
Solution Approach 2:
The patent applies parameter changes to normal product images by modifying image properties such as color values, brightness, contrast, and adding various defect patterns to transform them into synthetic defective images for training the vision inspection system
2Device complexity
If operator-dependent vision inspection methods are used, then device complexity is reduced, but measurement precision deteriorates due to variability in operator skill and fatigue
Solution Approach 1:
The patent replaces the mechanical human inspection process with an automated AI-based vision inspection system that uses trained neural networks to consistently identify defects, eliminating variability from operator skill and fatigue while maintaining operational efficiency
3Quantity of substance
If synthetic defective images are generated from normal product images, then the quantity of training data improves, but manufacturing precision deteriorates due to potential loss of real defect characteristics
Solution Approach 1:
The patent applies multiple diverse transformation operations (adding various defect types, color changes, rotations, cropping) to normal images to generate a large volume of synthetic defective images, using excessive transformation variety to compensate for any loss of realism in individual generated samples
Data Source
AI summary
An artificial intelligence device includes: a memory to store a first normal product image; a learning processor to train an image restoration model through inputting the first normal product image into the image restoration model as learning data to output a normal restored image similar to the first normal product image; and a processor configured to: modify the first normal product image to generate a first normal modified image belonging to a normal classification, and increase a number of a second normal product image belonging to the normal classification, modify at least one of the second normal product image to generate an abnormal modified image belonging to an abnormal classification, and input the abnormal modified image into the image restoration model to acquire an abnormal restored image output from the image restoration model.