Official code for paper: Counting melanocytes with trainable $h$-maxima and connected components counting layers accepted by DGMM2024 (Discrete Geometry and Mathematical Morphology):

Download the dataset from https://cloud.minesparis.psl.eu/index.php/s/c50xFQFENFZ6I5h

Install morpholayer:

cd DGMM2024_comptage_cellule; git clone https://github.com/Jacobiano/morpholayers.git

After changing --DATA_DIR in generate_preprocessed_numpy_dataset.sh to the dir containing database_melanocytes_trp1 :

bash generate_preprocessed_numpy_dataset.sh

This will create a directory named ./DGMM2024_comptage_cellule/best_h_dataset255, which contains the preprocessed numpy files and best h parameter ground truth.

After changing --DATA_DIR in test.sh to the dir containing database_melanocytes_trp1 :

bash test.sh

This will load the pretrained model weight and using the preprocessed inputs.

After testing finished, in the directory ./DGMM2024_comptage_cellule/visualize_test_only_hmaxima you can find the groud truth and detected data samples.

After changing --DATA_DIR in train.sh to the dir containing database_melanocytes_trp1 :

bash train.sh

This will train the CNN using preprocessed inputs from set1 in ./DGMM2024_comptage_cellule/, the best model weight with lowest validation error will be saved for each epoch.

If you find this code useful in your research, please consider citing:

@inproceedings{VelascoBMVC2022,
Author = {Velasco-Forero, S. and Rhim, A. and Angulo, J.},
Title = {Fixed Point Layers for Geodesic Morphological Operations},
Booktitle  = {British Machine Vision Conference (BMVC)},
Year = {2022}
}


@article{VelascoSIAM2022,
author = {Velasco-Forero, Santiago and Pag\`{e}s, R. and Angulo, Jesus},
title = {Learnable Empirical Mode Decomposition based on Mathematical Morphology},
journal = {SIAM Journal on Imaging Sciences},
volume = {15},
number = {1},
pages = {23-44},
year = {2022},
}