Figure 3

Hierarchical clustering, principal component analysis, and deep learning. Hierarchical clustering was performed for the mandible feature vectors, which consisted of 277 samples and 75 features (A). The darker color corresponds to the darker parts in the headshot image. Although genders were not separated effectively by this unsupervised method, female samples tended to exhibit darker parts around the halfway point. Principal component analysis was performed for the feature vectors (B). The blue and red dots represent 130 male and 147 female samples, respectively. Although they preferentially occupy the bottom left and top right corners, respectively, the two gender groups could not be clearly discerned by this unsupervised method. Learning curves of the CNN based on the Inception V3 is shown (C). Since the leave-one-out cross-validation with 10-times augmentation could achieve high accuracy in a single epoch, learning curves of a twofold cross-validation without data augmentation are shown. Odd-numbered and even-numbered volunteer samples were used as training and test data. Exemplary curves were observed during the first 5 epochs in this case. Cyan, pink, blue, and red lines indicate training loss, test loss, training accuracy, and test accuracy, respectively. The figure was drawn using Matplotlib 3.1.3 (https://matplotlib.org/) under Python 3.7.7 (https://www.python.org) and modified with Inkscape 1.1 (https://inkscape.org/).