Making automated image analysis simple with AI
While certain fields employ sophisticated AI techniques that are reproducible, robust, and validated, many metallography laboratories still suffer with methodologies that are time-consuming, lack reproducibility and robustness, and demand a high level of expertise in computer vision and/or AI from materials specialists.
For instance, methods such as point counting and manual intercept determination are laborious and often yield inaccurate or unreproducible results. A case is made with dendrite arm spacing measurements in aluminum castings.
Despite the industry's widespread adoption of threshold-based methods for automated image analysis, these techniques frequently falter due to their reliance on skilled computer vision professionals for development and their susceptibility to issues of reproducibility and accuracy from variations in illumination, sample preparation, and microstructure.
To address this issue, we propose two complementary approaches. Firstly, the creation of foundational models for image analysis that are robust to common variations in microscopy, sample preparation and generalize to most microstructure variations observed within a given domain. An example of this approach is grain size measurement in optical microscopy, where a foundational model has been developed and is presented herein.
Secondly, we emphasize the importance of addressing niche cases where the unique characteristics of the problem does not justify the large investment required for creating foundational models. In this case, we posit that the democratization of machine learning development through user-friendly, no-code tools with minimal learning curves is pivotal to fostering the adoption of automation in metallography laboratories. An example is provided by Clemex Studio, which shows the potential of this approach.