Similarity Measures for Satellite Images with Heterogeneous Contents

Abstract

Tversky's set-theoretic similarity states that a similarity measure should increase with the saliency of common features and decrease with that of distinctive features. When all necessary and relevant semantic features could be listed by hand, the similarity measure would be reduced to count the number of common features followed by subtracting the number of distinctive features. The reason is one might well select semantic features, so that features are independent and in the same level of salience. However, in image retrieval, one might have very restricted way to semantic features, for instance semantic features need to be derived from low-level features or by interaction between users and retrieval system. In this paper, we explore the Tversky's similarity measure between satellite images with heterogeneous contents in this situation. In this paper, the semantic feature is not any real word or phrase, but a label of class in which homogeneous regions reside. We assume that each region in images is related to a semantic feature, although we do not know what the semantic feature is. Therefore, semantic features used in this paper are not well defined in the sense that the distinction of different pair labels might vary from time to time. In other words, the salience of distinctive features might switch from one state to another. Therefore, a factor is proposed to simulate the Switch of Distinctive Features (SDF). The underlining principle is that the SDF would increase with the difference between two images in terms of content variation pattern within each image. Intuitively speaking, the role of distinctive features would be enlarged when there is little change in one image and clearly contrast in the other image. Although the definition of variation within single image is rather simple in this paper, experimental results show that the SDF does improve the retrieval precision of satellite images with heterogeneous contents.