色・形状・大きさを利用した物体概念の生成と評価

抄録

In general, CNN (Convolutional Neural Networks) is used as the method with high recognition accuracy. In CNN, however, several tens of thousands images are required as learning data for each category. Also, huge learning time is required. In contrast, after a human just look at several objects in a category he can get something like its general object concept. Furthermore, a human can represent the concept by words. In this article, a new concept learning method based on clustering and logistic regression is proposed, which requires low dimensional multi features and small training data. Generated object concept was evaluated in comparison with the result of recognition using real world objects included in RGB-D Object Dataset.