Synthetic Training Data Generation for Defect Detection Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision techniques face challenges in accurately detecting and localizing defects in production environments due to the high cost and inefficiency of manually labeling large datasets, leading to performance issues in deep learning models.
Innovation Solution
A method for generating synthetic training data using generative models like GANs and cGANs to create labeled datasets with balanced distributions, allowing for efficient training of data-driven models for defect detection and localization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large datasets with manual labeling are used for training deep learning models, then model performance improves, but time and cost increase significantly
Solution Approach 1:
The patent creates synthetic copies of real defect images through generative models (GANs, cGANs). These synthetic images replicate the visual characteristics and defect patterns of real images while being generated algorithmically, eliminating the need for manual labeling of additional real images. The synthetic training data preserves the essential features needed for model training without requiring time-consuming human annotation.
2Reliability
If more diverse training data is collected to improve detection accuracy, then model robustness improves, but data collection and labeling complexity increases
Solution Approach 1:
The patent systematically varies parameters in synthetic image generation including defect type, defect severity, object orientation, lighting conditions, and background variations. By controlling these parameters algorithmically, the system generates diverse training data covering multiple scenarios without the complexity of collecting and managing diverse real-world images. The conditional GAN takes defect type as input to generate specific defect variations.
Solution Approach 2:
The synthetic data generation system serves multiple functions: it generates diverse defect examples, balances class distributions, creates edge case scenarios, and provides unlimited training samples. A single generative model framework handles all these data requirements that would otherwise require multiple separate data collection efforts.
3Reliability
If balanced class distribution is achieved in training data, then model generalization improves, but data preparation effort increases
Solution Approach 1:
The generative model system automatically balances class distributions without manual intervention. The conditional GAN can generate equal numbers of samples for each defect class by controlling the input conditions, and can automatically create underrepresented classes. This self-balancing capability eliminates the need for manual data sampling, weighting, or re-sampling operations that would otherwise be required.
4Manufacturing precision
If realistic synthetic images are generated to maintain model accuracy, then image quality improves, but generation complexity increases
Solution Approach 1:
The patent uses an image-to-image translation model as an intermediary to transform synthetic images into more realistic appearances. The first GAN generates base synthetic images with correct defect characteristics, then the translation model refines these images to match the visual style and texture of real images. This two-stage approach maintains defect accuracy while improving visual realism without requiring direct manipulation of real images.
Data Source
AI summary
The present application relates to image processing. A computer-implemented method is provided for generating synthetic training data that is usable for training a data-driven model for analysing a surface image of a physical product that comprises at least one object, the method comprising:a) providing image data that comprises:an object image dataset comprising a plurality of object images of the at least one object, at least one object image being associated with a label usable for annotating a content of the object image; anda background image representing a background of a surface image of the physical product;b) generating a synthetic object image dataset from the object image dataset, wherein the synthetic object image dataset comprises a plurality of synthetic object images of the at least one object, at least one synthetic object image being associated with a label; andc) generating a plurality of first synthetic training data samples, wherein each first synthetic training data sample is generated by selecting one or more object images from the synthetic object image dataset and by plotting the selected one or more object images at one or more locations on the background image. The computer-implemented method may be used to improve the computer vision technique for the application in the technical field of agriculture and in production environment.