AI Defect Data Generation for Imbalanced Vision Inspection Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of securing a sufficient amount of high-quality defect data for training deep learning models in vision inspection systems is exacerbated by the imbalance between non-defect and defect data during the production process, making it difficult to achieve reliable defect detection.
Innovation Solution
An image processing apparatus generates virtual defect data using a Generative Adversarial Network (GAN) model to synthesize defect data based on non-defect data, enhancing the training of deep learning models by balancing the data imbalance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If defect data is collected during production process, then training data for deep learning model is obtained, but the number of defect data is limited and imbalance between non-defect data and defect data occurs
Solution Approach 1:
The patent uses image synthesis technology to create virtual defect images by copying and transforming real defect characteristics. The synthesis unit generates artificial defect images based on real defect data, allowing unlimited replication of defect types without physical constraints. This resolves the contradiction by providing abundant virtual defect data while maintaining the reliability of defect detection through accurate replication of real defect patterns.
Solution Approach 2:
The patent applies parameter transformation to convert real defect images into virtual defect images by changing various parameters such as defect position, size, shape, and visual characteristics. The transformation unit modifies defect parameters while preserving essential defect characteristics, enabling generation of diverse defect data from limited real samples. This increases the quantity of defect data without compromising detection reliability.
2Ease of manufacture
If deep learning model is trained with limited defect data, then model development is feasible, but the performance and accuracy of defect detection is insufficient
Solution Approach 1:
The patent creates virtual defect images that copy the essential characteristics of real defects, providing sufficient training data for model development. The synthesized images replicate defect patterns, positions, and visual features, enabling the deep learning model to learn accurate defect detection patterns from abundant virtual data while maintaining ease of training.
Solution Approach 2:
The patent performs preliminary synthesis of virtual defect images before model training. The synthesis unit pre-generates a comprehensive dataset of virtual defects, which is then used to train the deep learning model. This preliminary action ensures the model receives sufficient and diverse training data, improving detection accuracy while maintaining training feasibility.
3Reliability
If real defect data is used for training, then authentic defect characteristics are captured, but the data collection process is difficult and time-consuming
Solution Approach 1:
The patent creates virtual defect images by copying the essential characteristics of real defects through image synthesis. The synthesis unit replicates defect patterns, positions, and visual features from limited real samples, generating abundant virtual data that maintains authentic defect characteristics. This eliminates the time-consuming collection process while preserving the reliability of defect characteristic representation.
Solution Approach 2:
The patent transforms real defect images into virtual defect images by changing parameters such as position, size, shape, and visual properties. The transformation unit modifies defect parameters while preserving essential characteristics, enabling rapid generation of diverse training data without time-consuming physical collection. This maintains authenticity while dramatically reducing data collection time.
Data Source
AI summary
An image processing apparatus including a communication unit configured to receive at least one of training non-defect data and at least one of training defect data from an external device or a server, and a processor. The processor is configured to cause image degradation with respect to the training defect data according to a predetermined pattern or an arbitrary pattern, train an artificial intelligence generative model by using the degraded training defect data, extract defect information from a defect indicated by an image of a product, and generate final virtual defect data by inputting a second virtual defect data to the trained artificial intelligence generative model. In addition, the second virtual defect data is generated by synthesizing a first virtual defect data, a non-defect data and a first virtual mask image, and the first virtual defect data is generated based on the defect information.