Matching Words and Image Segments for Semantic Retrieval

概要

論文の詳細を見る
This paper presents a new method for retrieving images. A user's similarity interpretation is subjective and a similarity model's interpretation is objective. This method combines textual and object-based visual features to decrease this difference. It uses a novel multi-scale segmentation framework to detect prominent objects in an image. These objects are grouped depending on their visual features and mapped to related words obtained from psychophysical studies. Then, a hierarchy of words expressing higher-level meaning is determined on the basis of natural language processing and user evaluation. Experiments were carried out on 15,000 natural images. These showed higher retrieval precision in terms of estimating user retrieval semantics obtained via this two-layer association.
一般社団法人電子情報通信学会の論文
2005-06-09