Visualizing Attractive Factors in Buddhist Statue Images Using Grad-CAM
Hirofumi Shimoe, Hiroyuki Fujioka · 2023
In this study, we consider the problem of visualizing the attractive factors of Buddhist statue using a deep learning approach. For this purpose, we construct a Convolutional Neural Network (CNN) using the results of a questionnaire survey on the attractiveness of 553 Buddhist statue images conducted with 11 examinees as training data, and explore the attractive factors by visualizing the hidden layers of the CNN. For this visualization, we employ a method called as Gradient-weighted Class Activation Mapping (Grad-CAM), which combines feature maps and gradient information to represent important image regions contributing to the CNN output as heatmaps. The validity of the visualization results for these attractive factors is evaluated by experimental studies.