Classification of sand grains based on image data from electron microscopes

Tool on segmentation 5 sand grain microtextures (two chemical and three mechanical). The repository contains data, train/inference code and showcase of a final model. This repository was created as a part of batchelor thesis on the Faculty of Informational Technologies in Czech Technical University in Prague. Full research is available on DSpace

As a final segmentation model was selected DeepLabv3+. Model can segment microtextures with next IoU on test set:

Microtexture	IoU
pitting	0.111
edge_abrasion	0.197
precipitation	0.229
adhering_particles	0.415
conchoidal_fracture	0.452

How to run

The simplest way to run this tool is Google Colab Notebook.

If you want to run this project locally, make the next steps:

1. Clone this repository

git clone https://github.com/DainfromLiria/Sand_grains_classification.git

2. In the project folder run setup.sh script (firstly run chmod +x setup.sh if a script doesn't have execute permission). Next script makes python venv, install all requirements and merge model chunks into one model.pt file. Splitting on chunks allows loading a large model file on GitHub.

source setup.sh

3. Run jupyter notebook

jupyter notebook showcase.ipynb 

Train and validation

If you want to run model in train or validation mode, add .env file into the project directory. .env must contain Neptune project name and API key in format:

NEPTUNE_PROJECT=<project_name>
NEPTUNE_API_KEY=<api_key>

You can find this information inside your Neptune profile after you create a project.

Use run_train.sh for running the train loop.

Project Structure

• 📁 data - evaluation and train dataset with annotations in COCO format
• 📁 examples - images for README.md
• 📁 input_images - user images for prediction. Folder already contains six example images
• 📁 results - model weights and file with train settings
• 📁 scripts - utility scripts for visualization data from Neptun and dataset statistics. Script on split/merge 'model.pt'
• 📁 src - main project folder with source code
• 📄 run_train.sh - bash script for running model train loop using nohup
• 📄 setup.sh - bash script for setup venv, install requirements and merge model chunks into model.pt
• 📄 showcase.ipynb - showcase of final model for local run on PC.
• 📄 showcase_colab.ipynb - showcase of final model for Colab.