Automated Grain Size Analysis Using SAM

Quick Start

Installation

git clone https://github.com/ragh598/grain-size-analysis-sam.git
cd grain-size-analysis-sam

How It Works

1. Image Preprocessing: Load and resize microscopy images
2. Segmentation: Meta's SAM generates grain masks
3. Post-processing: Filter, clean, and validate masks
4. Morphometry: Calculate area, diameter, shape factors
5. Visualization: Generate annotated images and plots
6. Export: Save results to CSV with comprehensive statistics

Raw Image:

Model output:

High-throughput automated grain size analysis for materials science using Meta's Segment Anything Model (SAM)

Performance: Analyzed 10,830+ grains across multiple duralumin samples

• Mean grain area: 916.4 μm²
• Processing speed: 12 images/hous on GPU T4 on Google Colab
• This was implemented with sequential image input to the model. This code can be further optimized for multi-processing.
• ~90% reduction in analysis time vs. manual methods

Why This Project?

Manual grain size analysis in materials science is:

• Time-consuming: Hours per sample
• Subjective: Skill-dependent measurements
• Low-throughput: Limits statistical significance

This project aims to be a cost-effective alternative to manual and tedious checks for each image to check and identify the microstructure characteristics. This project automates the entire pipeline using SOTA Vision Transformer.