Foraminifera (forams for short) classification via deep feature extraction.
All models have been trained on a dataset of large, high-resolution images of
forams. The dataset has been produced by our research group, and will be made
publically available in the near future. Each of the source images consist of
a single class of forams. From these images, patches of 224x224 pixels are
extracted using combinations of Gaussian smoothing, binary image generation
via thresholding, and connected components. The first step removes the metallic
border present in all source images, and the second step extracts candidate
patches. Each patch that passes a defined selection critera is extracted by
placing a 224x224 crop at the centroid of the candidate region. The entire
process is automated in the preprocess_data.py
script.
Once the source images have been preprocessed by extracting patches, datasets
for training, validation, and testing are generated automatically by using the
build_datasets.py
script.
The raw-halves
source images are slightly different in nature, and requires
that the preprocess_data.py
script be invoked with --border-threshold=50
.
Patches from this dataset must be manually copied to the preprocessed
folder
built by process outlined above. In the future, we should find a way to fully
automate this step as well.