EFFICIENT SUPERVISED LEARNING METHOD WITH AUTOMATIC CLASSIFICATION OF TRAINING DATA SET

- USING GLOBAL AND STRUCTURAL GRAPHICAL FEATURES -

Abstract

Most of digital contents distributers use key words which correspond to objects in images to index various kinds of photo images. But these key words not always match with visual impression of images. We have developed a method to evaluate visual impression of images by using image key words which given by professional photographers. We have modeled their image evaluation process based on visual impressions (KANSEI Model). In this paper, we propose a method to efficient supervised learning. We succeed to improve accuracy of KANSEI Model by automatic classification of training data set and using two kinds of graphic features, global and structural graphic features which imitated visual neuron of human. By using the KANSEI model, we have developed automatic image classification system for various kinds of photo images based on visual impressions, and we have applied to similar image search system.