In query-by-semantic-example image retrieval, images are ranked by similarity of semantic descriptors. These descriptors are obtained by classifying each image with respect to a pre-defined vocabulary of semantic concepts. In this work, we consider the problem of improving the accuracy of semantic descriptors through cross-modal regularization, based on auxiliary text. A cross-modal regularizer, composed of three steps, is proposed. Training images and text are first mapped to a common semantic space. A regularization operator is then learned for each concept in the semantic vocabulary. This is an operator which maps the semantic descriptors of images labeled with that concept to the descriptors of the associated texts. A convex formulation of the learning problem is introduced, enabling the efficient computation of concept-specific regularization operators. The third step is the selection of the most suitable operator for the image to regularize. This is implemented through a quantization of the semantic space, where a regularization operator is associated with each quantization cell. Overall, the proposed regularizer is a non-linear mapping, implemented as a piecewise linear transformation of the semantic image descriptors to regularize. This transformation is a form of cross-modal domain adaptation. It is shown to achieve better performance than recent proposals in the domain adaptation literature, while requiring much simpler optimization. © 2014 Elsevier Inc. All rights reserved.

The problem of cross-modal retrieval from multimedia repositories is considered. This problem addresses the design of retrieval systems that support queries across content modalities, for example, using an image to search for texts. A mathematical formulation is proposed, equating the design of cross-modal retrieval systems to that of isomorphic feature spaces for different content modalities. Two hypotheses are then investigated regarding the fundamental attributes of these spaces. The first is that low-level cross-modal correlations should be accounted for. The second is that the space should enable semantic abstraction. Three new solutions to the cross-modal retrieval problem are then derived from these hypotheses: correlation matching (CM), an unsupervised method which models cross-modal correlations, semantic matching (SM), a supervised technique that relies on semantic representation, and semantic correlation matching (SCM), which combines both. An extensive evaluation of retrieval performance is conducted to test the validity of the hypotheses. All approaches are shown successful for text retrieval in response to image queries and vice versa. It is concluded that both hypotheses hold, in a complementary form, although evidence in favor of the abstraction hypothesis is stronger than that for correlation.