Bolete Classifier

Bolete identification algorithm that utilizes modern deep learning techniques and domain knowledge about important mushroom characteristics.

Paper

The preprint is available in PDF form in Bolete_Species_and_Edibility_Classification.pdf in the root of the project.

Prerequisites

Codebase was created and has only been tested on Ubuntu 20.04.

Using a virtual env or conda environment, run

pip install -r requirements.txt

This will install the python dependencies.

For the data preprocessing pipeline, install MATLAB R2020b.

Pipeline

To run the pipeline from the raw dataset, which is unavailable to the public at this time:

cd data_pipeline/
sh run_pipeline.sh

This will:

1. Get the list of possible categories of mushroom for the system to predict from the tabular data files.
2. Collect valid labels for the data by cross referencing which mushrooms have (or do not have) those categories in the tabular data.
3. Copy the images from the original photos directory by the name that the photos are titled, for example Allesioporus-rubriflavus-Jane-Smith.jpeg'' would be copied to the folder allesioporus-rubriflavus''.
4. Copy the images referenced by the tabular data that correspond to our mushroom labels into the appropriate folder and copy non-duplicates from the folder created from filenames directly in step 3. Now clean_data/labels.csv contains all the mushrooms and the mushroom identification features to predict. And the data folder contains photos for each mushroom by the newest name of the mushroom for data-augmentation purposes (add correct photos to these bins as needed).
5. Process all image files in data by resizing and padding them so they are 512 x 512. Write these compressed versions to an HDF5 dataset file with their corresponding labels.

Building

To build the Cython files execute

python setup.py build_ext --inplace

Running Models

The direct model variants can be run from the direct.ipynb file and the characteristic model variants can be run using the characteristic.ipynb.

Options:

• FINAL = True if you want to produce evaluation results on the test set, False otherwise
• EDIBILITY = True if the direct model should output edibility scores, False for species scores
• HIGH_RES = True to use 512x512 images, False to use 256x256 images.

Evaluating Results

Running the cells in performance.ipynb will produce the performance data from the data in evaluation_data/. Output will be located in performance_data/

To generate all the plots used in the paper, execute:

>> paired_history_plots
>> p_charts

Acknowledgements

Data generously provided by the Western Pennsylvania Mushroom Club via collaboration with Scott Pavelle and Richard Jacob.