Skip to main content
Fig. 3 | Heritage Science

Fig. 3

From: Discerning the painter’s hand: machine learning on surface topography

Fig. 3

Patch size versus two measures of ML performance: overall accuracy (thick black curve) and F1 scores for each artist (thin colored curves). The best results occur for patches in the range 100–300 pixels (~ 5–15 mm). The overall accuracy decreases when patch size is smaller than 100 pixels (due to lack of information in each patch) and when patch size is larger than 300 pixels (because the size of the training data set decreases with increasing patch size). Error bars are standard deviations over different repetitions of ensemble training. For comparison, the maximum likelihood estimation (MLE) accuracy results based on the pixel height distributions are shown as a dashed curve

Back to article page