Sampyo & SNU : SampyoNet

Convolutional neural network developed by Seoul National University for Sampyo Cement to facilitate an automated assessment of gravel quality.

Instructions

1. Using Python 3, download all the requirements via requirements.txt.

   python3 -m pip install -r requirements.txt

2. Run the following command to predict on single image

   python3 main.py -w models/[WEIGHTS_FILE].hdf5 -i [IMAGE_FILE].jpg

3. Run the following command to predict on a set of images specified through a CSV file

   python3 main.py -w models/[WEIGHTS_FILE].hdf5 -a [CSV_FILE].csv

4. Implement into applications using the wrapper function predict_image in main.py

5. To further train the model, TensorFlow training script must be added. This is just a few more lines of code which can be written very easily. The code is not included here for simplicity.

Project Structure

main.py contains the wrapper functions for running the model and can be initiated directly from the command line.

model.py defines the model schema.

config.py has all the required import statements and global parameters which should not be modified.

brightness.py contains the functions which calculate the brightness of the input image.

requirements.txt lists all the dependencies.

models/ is a directory with all pretrained weights in it. Our main model is 2_largfac.hdf5

preprocess/ is a directory with sources used in preprocessing stage.

Reference

This model was inspired by SediNet, a deep learning convolutional neural network made by Dr. Daniel Buscombe (github repo: https://github.com/MARDAScience/SediNet)