SYNOSIS: Image Synthesis Pipeline for Machine Vision in Metal Surface Inspection | JoVE Visualize

Area of Science:

Computer Vision
Machine Learning
Materials Science

Background:

Machine learning (ML) for visual inspection requires extensive, diverse training data, which is often impractical to acquire.
Existing synthetic data generation methods, primarily generative models, face challenges like data shortages, hallucinations, and limited handling of edge-cases.
A gap exists in generating physically realistic textures with structured patterns for robust ML model training.

Purpose of the Study:

To present a novel synthetic data generation pipeline for creating large datasets of physically realistic textures with structured patterns.
To enable precise control over texture parameters for generating diverse observed and unobserved texture instances.
To evaluate the quality of the generated synthetic dataset and its potential for predicting downstream ML performance.

Main Methods:

Development of a synthetic data generation pipeline based on procedural texture modeling with interpretable parameters.
Generation of datasets featuring sandblasting, parallel milling, and spiral milling textures.
Evaluation of dataset quality through image similarity metrics between real and synthetic domains, beyond final model performance.

Main Results:

The pipeline successfully generated large datasets of physically realistic textures with sophisticated structured patterns.
The procedural approach allowed for precise control over texture parameters, ensuring variety and coverage of edge-cases.
Image similarity metrics indicated a trend that could predict downstream detection performance, offering a new evaluation method.

Conclusions:

The proposed pipeline offers a robust solution for generating high-quality synthetic texture data, addressing limitations of current methods.
The interpretable parameter control provides unique advantages for creating diverse and realistic datasets for machine learning.
The findings suggest that image similarity metrics can serve as valuable predictors for downstream task performance, guiding future synthetic data development.