Geoscience document layout for figures and figure classification into 48 geoscience image categories
Augmented and non augmented training images for the yolov5s DLA (document layout) model can be freely downloaded and be used at https://universe.roboflow.com/peter-j1jzx/findthe-images. The pretrained DLA model is in the repro.
In this repro you can find the code to extract geoscience figures from typical oil and gas well reports and other typical oil and gas documents like conference papers or relinquishment reports
Additionally we provide a pretrained tensorflow Resnet50 V2 model trained on Microsoft LOBE https://www.lobe.ai/ for the classification of the extracted geoscience figures. The figure classifier has been trained on 30.0000 publicly available geoscience images extracted from a variety of data soruces (open access journals, relinquishment reports, NPD wellbore documents, NLOG publications.
48 image categories that cover exploration and production geoscience have been labeled and trained on. The published model is a good starting point to build your own more refined figure classifier.
These are the categories
(_
3d_block_diagram,
3d_display,
background_logos,
backgrounds,
basin_mod_burial_curve,
basin_mod_chromatogram,
basin_mod_maturity_depth,
basin_mod_van_krevelen_diagram,
core_description_log,
core_images,
cuttings images,
cuttings images uv,
geosection,
graph_fault_seal,
graph_histogram,
graph_plot_xy,
graph_rose_diagram,
graph_ternary plot,
installations,
map_contour,
map_culture,
map_legend,
map_petrol_system,
map_risk_crs,
map_seismic_attribute,
map_seismic_depth,
outcrop_images,
people_images,
seismic_avo_model,
seismic_gathers,
seismic_inversion,
seismic_lines,
seismic_polarity,
seismic_synthethic_seismogram,
seismic_wedge_model,
seismic_with_well,
stratigraphic_table,
table_classic,
thin_section_sem_images,
thin_section_through_light,
volumetrics_decision_tree,
volumetrics_geox_graphs,
volumetrics_geox_tornado_diagram,
well_log,
well_log_correlation,
well_log_fmi_images,
well_pressure_prediction
well_schematic),_
The training images for the yolov5s DLA (document layout) model can be freely downloaded and be used at https://universe.roboflow.com/peter-j1jzx/findthe-images. You can also export the images cropped to the bounding boxes and use this for classification rather than segmentation
