Fig. 1
From: Classifying polish in use-wear analysis with convolutional neural networks
Summary of final accuracy and loss values reached by the models combining different architectures, patch sizes and objectives. Color gradients distinguish model configurations (upper legend): custom CNNs (red-purple) and ResNet50 (blue-green), with hues varying by patch size (9 vs. 16) and objective (10× vs. 20×), (lower legend): custom CNN (red) and ResNet50 (blue). (a–d) Scatterplots comparing metrics across training and validation sets: (a) training loss vs. training accuracy, (b) validation loss vs. validation accuracy, (c) training loss vs. validation loss, and (d) training accuracy vs. validation accuracy. (e–h) Kernel density estimates showing distributions of each metric per model type (e) training accuracy, (f) validation accuracy, (g) training loss, and (h) validation loss.
