Abstract-This paper proposes a tree-based pursuit algorithm that efficiently trades off complexity and approximation performance for overcomplete signal expansions. Finding the sparsest representation of a signal using a redundant dictionary is, in general, an NP-hard problem. Even suboptimal algorithms such as Matching Pursuit remain highly complex. We propose a structuring strategy that can be applied to any redundant set of functions, and which basically groups similar atoms together. A measure of similarity based on coherence allows for representing a highly redundant subdictionary of atoms by a unique element, called molecule. When the clustering is applied recursively on atoms and then on molecules, it naturally leads to the creation of a tree structure. We then present a new pursuit algorithm that uses the structure created by clustering as a decision tree. This tree-based algorithm offers important complexity reduction with respect to Matching Pursuit, as it prunes important parts of the dictionary when traversing the tree. Recent results on incoherent dictionaries are extended to molecules, while the true highly redundant nature of the dictionary stays hidden by the tree structure. We then derive recovery conditions on the structured dictionary, under which tree-based pursuit is guaranteed to converge. Experimental results finally show that the gain in complexity offered by tree-based pursuit does in general not have a high penalty on the approximation performance. They show that the dimensionality of the problem is reduced thanks to the tree construction, without significant loss of information.

The performances of approximation using redundant expansions rely on having dictionaries adapted to the signals. In natural high-dimensional data, the statistical dependencies are, most of the time, not obvious. Learning fundamental patterns is an alternative to analytical design of bases and is nowadays a popular problem in the field of approximation theory. In many situations, the basis elements are shift invariant, thus the learning should try to find the best matching filters. We present a new algorithm for learning iteratively generating functions that can be translated at all positions in the signal to generate a highly redundant dictionary.

Abstract-Real-world phenomena involve complex interactions between multiple signal modalities. As a consequence, humans are used to integrate at each instant perceptions from all their senses in order to enrich their understanding of the surrounding world. This paradigm can be also extremely useful in many signal processing and computer vision problems involving mutually related signals. The simultaneous processing of multimodal data can, in fact, reveal information that is otherwise hidden when considering the signals independently. However, in natural multimodal signals, the statistical dependencies between modalities are in general not obvious. Learning fundamental multimodal patterns could offer deep insight into the structure of such signals. In this paper, we present a novel model of multimodal signals based on their sparse decomposition over a dictionary of multimodal structures. An algorithm for iteratively learning multimodal generating functions that can be shifted at all positions in the signal is proposed, as well. The learning is defined in such a way that it can be accomplished by iteratively solving a generalized eigenvector problem, which makes the algorithm fast, flexible, and free of user-defined parameters. The proposed algorithm is applied to audiovisual sequences and it is able to discover underlying structures in the data. The detection of such audio-video patterns in audiovisual clips allows to effectively localize the sound source on the video in presence of substantial acoustic and visual distractors, outperforming state-of-the-art audiovisual localization algorithms.

No abstract