Optimizing Image Segmentation for Microstructure Analysis of High-Strength Steel: Histogram-Based Recognition of...

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Accurate analysis of material microstructures is essential for understanding material properties.
• Unsupervised image segmentation is a key preprocessing step for microstructure analysis.
• The performance of classification tasks is often dependent on the quality of segmentation.

Purpose of the Study:

• To systematically compare the performance of three unsupervised segmentation algorithms (SLIC, Felzenszwalb, Watershed) when combined with two classification approaches (Random Forest, CNNs).
• To investigate the impact of segmentation parameter tuning on downstream classification performance using Bayesian optimization.
• To evaluate the trade-offs between feature-engineered and end-to-end learning methods for microstructure analysis.

Main Methods:

• Comparison of Simple Linear Iterative Clustering (SLIC), Felzenszwalb's graph-based method, and Watershed algorithms.
• Integration of segmentation algorithms with Random Forest (using histogram features) and Convolutional Neural Networks (CNNs).
• Bayesian optimization for joint tuning of segmentation parameters and model hyperparameters.
• Validation using light optical microscopy images of high-strength steel, assessed via stratified cross-validation and independent test sets.

Main Results:

• Significant impact of segmentation algorithm choice on classification performance was observed.
• Bayesian optimization effectively tuned parameters for improved segmentation and classification.
• Convolutional Neural Networks (CNNs) showed competitive performance, especially with optimized segmentation.
• The study highlights the critical role of segmentation in microstructure image analysis.

Conclusions:

• Segmentation algorithm selection is a critical factor in the success of automated microstructure analysis.
• The choice between feature-engineered and end-to-end learning approaches depends on specific analysis goals and data characteristics.
• Optimized segmentation parameters enhance the reliability of microstructure classification for materials science applications.