# Optimizing Image Segmentation for Microstructure Analysis of High-Strength Steel: Histogram-Based Recognition of Martensite and Bainite

## TL;DR

This paper compares image segmentation methods for analyzing steel microstructures, showing how segmentation choices affect classification accuracy.

## Contribution

A systematic comparison of segmentation and classification methods for steel microstructures using Bayesian optimization.

## Key findings

- SLIC and Felzenszwalb's method outperformed the Watershed algorithm for segmentation.
- Random Forest with histogram features performed competitively against CNNs.
- Segmentation quality significantly impacts downstream classification performance.

## Abstract

This study systematically compares three unsupervised segmentation algorithms (Simple Linear Iterative Clustering (SLIC), Felzenszwalb’s graph-based method, and the Watershed algorithm) in combination with two classification approaches: Random Forest using histogram-based features and Convolutional Neural Networks (CNNs). The study employs Bayesian optimization to jointly tune segmentation parameters and model hyperparameters, investigating how segmentation quality impacts downstream classification performance. The methodology is validated using light optical microscopy images of a high-strength steel sample, with performance evaluated through stratified cross-validation and independent test sets. The findings demonstrate the critical importance of segmentation algorithm selection and provide insights into the trade-offs between feature-engineered and end-to-end learning approaches for microstructure analysis.