Abstract: In this paper we perform statistical data analysis of a broad set of state-of-the-art audio features and low-level MPEG-7 audio descriptors. The investigation comprises data analysis to reveal redundancies between state-of-the-art audio features and MPEG-7 audio descriptors. We introduce a novel measure to evaluate the information content of a descriptor in terms of variance. Statistical data analysis reveals the amount of variance contained in a feature. It enables identification of independent and redundant features. This approach assists in efficient selection of orthogonal features for content-based retrieval. We believe that a good feature should provide descriptions with high variance for the underlying data. Combinations of features should consist of decorrelated features in order to increase expressiveness of the descriptions. Although MPEG-7 is a popular and widely used standard for multimedia description, only few investigations do exist that address analysis of the data quality of lowlevel MPEG-7 descriptions.

In this paper we perform statistical data analysis of a broad set of state-of-the-art audio features and low-level MPEG-7 audio descriptors. The investigation comprises data analysis to reveal redundancies between state-of-the-art audio features and MPEG-7 audio descriptors. We introduce a novel measure to evaluate the information content of a descriptor in terms of variance. Statistical data analysis reveals the amount of variance contained in a feature. It enables identification of independent and redundant features. This approach assists in efficient selection of orthogonal features for content-based retrieval. We believe that a good feature should provide descriptions with high variance for the underlying data. Combinations of features should consist of decorrelated features in order to increase expressiveness of the descriptions. Although MPEG-7 is a popular and widely used standard for multimedia description, only few investigations do exist that address analysis of the data quality of low-level MPEG-7 descriptions. 1

The main objective of this work is to study and implement techniques for visual content retrieval using relevance feedback. Relevance feedback approaches make use of interactive learning in order to modify and adapt system behaviour to user’s desires by modelling human subjectivity. They allow a more semantic approach based on user’s feedback information, while relying on similarity derived from low-level features. An image relevance feedback framework has been implemented based on support vector machines as a generalisation method. The algorithm for support vector machines solves a convex optimisation problem and the algorithm has been tailored to the relevance feedback scenario. MPEG-7 standard descriptors and their recommended distance functions have been used to represent low-level visual features as well as several additional descriptors. A multi-feature scenario has been developed in an effort to represent visual content as close as possible to human perceptual experience. A model for feature combination and not just concatenation has been developed and a novel kernel for adaptive similarity matching in support vector machines has been proposed. The new kernel models multi-feature space guaranteeing convergence of the support vector optimisation problem. To address the problem of visual content representations, a novel approach of building descriptors based on image blocks, their low-level features and their spatial correlation has been proposed as a part of the relevance feedback framework. In accordance to this an accompanying kernel on sets has been proposed that handles both multi-feature space as well as the local spatial information among image blocks. The relevance feedback module has been applied to a framework for image selection in concept learning. It combines unsupervised learning to organize images based on low-level similarity, and reinforcement learning based on relevance feedback, to refine the

Project full title Bootstrapping Ontology Evolution with Multimedia Information Ex-traction

This research work addresses the problem of using concept-related indexing of image content as a near-automatic way to perform semantic image annotation. The main objective is to pro-vide a framework in which lexical information of visual interpretations and their components (concept-related indexes) can be used to perform content-based image annotation. Several design phases starting from the formation of an MPEG-7 learning space to the construction of a robust semantic indexer were applied. Salient features of the proposed framework for concept-related indexing of image content are: Provide a suitable combination of low-level visual features. A specific concern is on defining a structure in which MPEG-7 descriptor elements may be aggregated into feature vectors. A structure is proposed to preserve the semantics embedded in descriptors, avoid description overriding, and control the vector dimensionality using the minimum number of required elements. Unambiguous interpretation is reduced using a built-in knowledge base consisting of con-cepts organized into a restrained lexicon.

Abstract not found