Enhancing Quantification of Inclusions in PoDFA Micrographs Through Integration of Deterministic and Deep Learning Image Analysis Algorithms

Abstract

The assessment of aluminum melt cleanliness has traditionally relied on labor-intensive and subjective manual processes. The present study builds upon prior digital image analysis to quantify inclusions in micrographs of PoDFA samples. Through the integration of deterministic methods, unsupervised Machine Learning (ML), and neural networks, cleanliness data comparable to PoDFA grid assessments has been achieved. Overcoming the challenge of generating sufficient and accurate training data for neural networks, the suggested approach has been refined. Enhanced isolation strategies for target classes have resulted in higher-quality training data, elevating the prediction accuracy of the neural network. Post-processing of neural network predictions has also been improved. The integrated approach presented here demonstrates more reliable cleanliness data than previous implementations. Offering a promising alternative to manual PoDFA assessments, this integrated approach improves efficiency and reduces human biases.