Visual Inspection Training Data Using Synthetic Defect Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional visual inspection methods in pharmaceutical production face challenges in generating sufficient and high-quality training data for machine learning models, leading to high false reject rates and inadequate representation of defect properties, which is crucial for GMP compliance.
Innovation Solution
A method of generating training data by combining defect images with good product images to create a wide range of defect characteristics, using techniques like image segmentation and generative adversarial networks (GANs) to enhance the training dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If artificial defects are created by introducing foreign particles or deliberately damaging products, then training data for defective products can be obtained, but the process becomes time-consuming and costly
Solution Approach 1:
The patent uses generative adversarial networks (GANs) to create synthetic defect images that copy and replicate real defect characteristics without requiring physical manipulation of products. The GAN generates realistic defect variations by learning from a small set of real defect images, thereby producing large quantities of training data without the time-consuming process of manually creating artificial defects through foreign particle introduction or deliberate damage.
Solution Approach 2:
The patent replaces the mechanical process of physically creating defects (introducing foreign particles, deliberate damage) with a computational image generation system. Instead of mechanically altering physical products to create training samples, the system uses neural networks to synthesize defect images, substituting mechanical defect creation methods with automated digital generation processes that are faster and more cost-effective.
2Reliability
If the full range of defect variations is mapped in training data, then classification accuracy improves, but the complexity and cost of creating training data increases significantly
Solution Approach 1:
The patent employs a dynamic training data generation approach where the GAN continuously learns and adapts to generate diverse defect variations. The system dynamically adjusts the complexity and variety of generated defects based on the training progress and performance requirements, allowing comprehensive coverage of defect space without manually designing each variation. This dynamic generation process simplifies the overall complexity by automating the exploration of defect variations.
Solution Approach 2:
The patent utilizes parameter changes in the GAN model to generate diverse defect variations by modifying parameters such as defect size, shape, position, and type during the image generation process. By systematically varying these parameters, the system comprehensively maps the full range of defect variations needed for accurate classification, achieving high reliability without the complexity of manually creating each variation scenario.
3Reliability
If actual defects from production processes are used to create training data, then real defect representation is achieved, but the lack of sufficient real defects makes this approach insufficient
Solution Approach 1:
The patent uses GANs to copy and replicate the characteristics of real defects found in production processes. By training the GAN on a small set of actual defect images, the system learns to generate numerous synthetic defect images that faithfully reproduce real defect patterns, thereby overcoming the limitation of having insufficient real defect samples while maintaining authentic defect representation.
Solution Approach 2:
The patent performs preliminary action by collecting and analyzing real defect characteristics from production processes before generating the full training dataset. The GAN is pre-trained on authentic defect samples to learn real defect patterns, and then uses this knowledge to generate comprehensive training data. This preliminary analysis of real defects ensures that the subsequently generated synthetic defects accurately represent actual production defects, achieving both quality representation and sufficient quantity.
Data Source
AI summary
A method of generating training data for training a machine learning model for visual product inspection is provided, wherein the training data are generated to comprise a plurality of first product images associated with a first class and a plurality of second product images associated with a second class. The method comprises obtaining a plurality of defect images, each representing at least one defect that can occur in the product, wherein the plurality of defect images is obtained in such a way to represent a plurality of different defects that can occur in the product, and creating a plurality of combined images. The creating of the plurality of combined images comprises obtaining a product image representing a product without a defect, combining the product image with at least one defect image of the plurality of defect images to obtain a combined image representing the product with at least one defect, and associating the obtained combined image with the second class. The plurality of combined images is added to the plurality of second product images.