Fig 1.
Photomicrographs (left) and labels (right).
Table 1.
Number of images in the train, validation and test sets for each class.
Table 2.
Relative size of regions compared to the image size and number of regions per image.
Fig 2.
Segmentation and classification workflow.
Table 3.
Model performance in the bone class.
Table 4.
Model performance in the charcoal class.
Table 5.
Model performance in the flint/obsidian class.
Fig 3.
Bone class photomicrograph, ground truth, prediction and comparison of the segmentation, where false positives are shown in green and false negatives in red.
Fig 4.
Charcoal class photomicrograph, ground truth, prediction and comparison of the segmentation, where false positives are shown in green and false negatives in red.
Fig 5.
Flint/obsidian class photomicrograph, ground truth, prediction and comparison of the segmentation, where false positives are shown in green and false negatives in red.
Table 6.
Number and percentage of images per abundance level.
Fig 6.
Confusion matrices of relative abundance classification.
Confusion matrices obtained for the classification of relative abundance using the U-Net model with InceptionV4 as encoder in the test sets for flint/obsidian, charcoal and bone. The categories are: Very Few (<5%), Few (6–15%), Common (16–30%), Frequent (31–50%), Dominant (51–70%), and Very Dominant (>70%).
Table 7.
Classification performance across abundance categories using U-Net with InceptionV4.