Prosemantic Features for Content-Based Image Retrieval

Abstract

Esta es la versión de autor de la comunicación de congreso publicada en: This is an author produced version of a paper published in: Abstract. We present here, an image description approach based on prosemantic features. The images are represented by a set of low-level features related to their structure and color distribution. Those descriptions are fed to a battery of image classifiers trained to evaluate the membership of the images with respect to a set of 14 overlapping classes. Packing together the scores vectors of prosemantic features are obtained, and used to index the images in an image retrieval system. To verify the effectiveness of the approach, we designed a target search experiment in which both low-level and prosemantic features are embedded into a content-based image retrieval system exploiting relevance feedback. The experiments show that the use of prosemantic features allows for a more successful and quick retrieval of the query images.

Spatially organized visualization of image query results

“…In the experimentation reported in our previous works 7,8 prosemantic features demonstrated to perform significantly better for image retrieval than low-level features. In our opinion these performance derives from the capability of prosemantic features of allowing a better match against users' intuition about the similarity of the images.…”

“…These features, which we called prosemantic (from "towards the meaning"), 7,8 are compact 56-dimensional feature vectors. The small dimensionality of the feature space allows for an efficient analysis of the distribution of the images.…”

“…We originally introduced prosemantic features in a previous work 7 and we further explored them in another paper.…”

Unsupervised Classemes

“…Ciocca et al presented an image descriptor, that they called "prosemantic features", based on the output of a number of image classifiers [3]. The feature vector is created, by concatenating the output of 56 different soft classifiers trained to identify 14 different classes on the basis of four different low-level features.…”