Image Feature Evaluation for Contents-based Image Retrieval

Kuffner, A. and Robles-Kelly, A.

This paper is concerned with feature evaluation for content-based image retrieval. Here we concentrate our attention on the evaluation of image features amongst three alternatives, namely the Harris corners, the maximally stable extremal regions and the scale invariant feature transform. To evaluate these image features in a content-based image retrieval setting, we have used the KD-tree algorithm. We use the KD-tree algorithm to match those features corresponding to the query image with those recovered from the images in the data set under study. With the matches at hand, we use a nearest neighbour approach to threshold the Euclidean distances between pairs of corresponding features. In this way, the retrieval is such that those features whose pairwise distances are small, 'vote' for a retrieval candidate in the data-set. This voting scheme allows us to arrange the images in the data set in order of relevance and permits the recovery of measures of performance for each of the three alternatives. In our experiments, we focus in the evaluation of the effects of scaling and rotation in the retrieval performance.

Cite as: Kuffner, A. and Robles-Kelly, A. (2006). Image Feature Evaluation for Contents-based Image Retrieval. In Proc. HCSNet Workshop on the Use of Vision in Human-Computer Interaction, (VisHCI 2006), Canberra, Australia. CRPIT, 56. Goecke, R., Robles-Kelly, A. and Caelli, T., Eds. ACS. 29-33.

(from crpit.com) (local if available)