Comparing transfer learning to feature optimization in microstructure classification.

Area of Science:

• Materials Science
• Machine Learning
• Computational Biology

Background:

• Manual analysis of research data, particularly microstructural morphologies, is time-consuming and inefficient.
• Machine learning (ML) has shown promise in automating tasks like image segmentation and classification.
• Existing ML models, including transfer learning with VGG16, InceptionV3, and Xception, provide moderate performance for binary classification of microstructures.

Purpose of the Study:

• To improve binary classification models for high-throughput identification of microstructural morphologies.
• To investigate the efficacy of feature engineering combined with a novel algorithm compared to transfer learning alone.
• To enhance the classification accuracy for datasets with limited observations.

Main Methods:

• Utilized a dataset of 133 dendritic and 444 non-dendritic structures.
• Applied data augmentation (rotation, translation) to increase dataset size six-fold.
• Employed transfer learning with pre-trained networks (VGG16, InceptionV3, Xception) and a new nature-inspired algorithm, the Binary Red Deer Algorithm (BRDA), for feature optimization and classification.

Main Results:

• Transfer learning methods achieved moderate F1 scores ranging from 0.801 to 0.822.
• The Binary Red Deer Algorithm (BRDA) achieved significantly higher F1 scores in the range of 0.96.
• Feature engineering via BRDA demonstrated superior performance over transfer learning alone for this classification task.

Conclusions:

• The Binary Red Deer Algorithm (BRDA) offers a more effective approach for binary classification of microstructural morphologies compared to standard transfer learning techniques.
• Feature engineering plays a crucial role in enhancing classification accuracy, especially with limited datasets.
• This work highlights the potential of nature-inspired algorithms for advancing high-throughput data analysis in scientific research.