Abstract
The construction of training datasets for machine learning to estimate the image impression of generally requires annotating a large number of images. Here, it is problematic in that the training results depend on the individual impression responses of the operators. In this study, we propose a semi-automatic system for annotating impressions of images based on the results of machine learning of the impression responses of many operators, and a method for visualizing the annotation process in order to reduce the individual differences in the impressions of images and the burden on the operator. First, the system evaluates the impressions using the SD method, and then generates a fuzzy decision tree using the impression values of each image. The fuzzy decision tree automatically classifies images, and then visualizes the results and the process of classification to assist the user in reclassifying images. By linking the display of the decision tree with a similarily-based image browser, we improve the readability of the decision tree and support understanding of annotation trends. In this paper, we present an example of visualization using this method on data from 43 workers' impression evaluations of 1,500 images of women's clothing, and verify the usefulness of this research through user experiments.
