Generation of Stratified Image Database with Web Image Sharing Service and Ontology

Abstract

Preparation of a training image dataset used for the genetic object recognition system needs a lot of costs. Some datasets were recently created from Web image sharing services like Flickr, but examples which utilize a semantic classification or information of tags given to photos are quite few. This paper proposes an automatic generation method of an image dataset by introducing semantics of each label given to images, where the images are positioned in a semantic stratification according to the meaning of labels. Here the semantic stratification consists of IS-A relations, where "Animal" locates in a upper concept of "Dog" and "Cat". By using an image dataset with conceptual information like this, a computer will be able to recognize objects pictured in photos from a wide concept like "Animal" or "Vehicle" to narrower concepts gradually. In experiments, 47,910 photos were automatically classified into 183 classes which were aligned in a stratified tree. An evaluation test of the labels was performed manually for the generated image dataset and it was confirmed that most of labeling were reasonably collected.