Resumo:
This project presents an iterative approach for upscaling a machine learning model for microstructural semantic segmentation of two-phase steels light optical micrographs. Several deep learning models have been trained, using a U-NET architecture with DenseNet-201 pretrained weights on ImageNet as backbone.
Metallographic samples from rolled plates have been produced and analyzed in different microscopes to collect data for training and testing, aiming to specifically increase the manageable variance as well as the model’s robustness. The results from previous models were then used as masks to train the final one.
The incorporation of a higher variance in the model through different acquisition conditions images resulted in a more robust model, that can consistently segment images at various magnifications, from different microscopes, and taken under suboptimal conditions.
The utilization of previous segmentation results as masks allowed to introduce more data to the training data set. This allowed to minimize the need to produce hand annotated masks, which are time consuming to make and often constitute a bottle neck for model training.
The relevance of these results lies in the possibility to correlate the results from the model (second phase fraction and morphological parameters of the particles) with mechanical properties and manufacturing parameters. Moreover, light optical micrographs are inexpensive, fast to produce and already implemented in quality control at an industrial scale, thus making the implementation of this analysis technique in the industry feasible.
