出版社：The Editorial Committee of the Interdisciplinary Information Sciences
摘要：With advances in digital technologies, the number of images we are subjected to every day has increased significantly. Predicting and recommending human subjective preferences for images is useful for selecting image data efficiently to avoid the unnecessary use of valuable storage space. In this study, we investigate the use of a machine learning model for estimating human preferences for images from spontaneous facial features extracted from video images of human faces while they are performing a natural preference evaluation task. We use two image categories and compare the results between categories. We also conduct an experiment to assess the performance of human raters in predicting the preferences of others from facial videos. As a standard to compare predictive performance from facial expressions, we also test prediction from high-level image features by training a deep learning model using the obtained experimental data. The results show that the spontaneous facial features produce prediction performance comparable with, and for lunch box images, marginally better than, the image features specifically trained for our dataset, and clearly outperform the human raters. We further examine which facial expression features are important for prediction and show that the important facial features differ between image categories. Our results show that facial expressions can be used to predict the preference for images, to some extent, although we need to be careful when generalizing the learned model to other image categories. Our machine learning approach also provides insights into the differences in the cognitive mechanisms used for preference evaluation for different image categories.
