Due to the rapid development of job markets, traditional recruitment methods are becoming insufficient. This is because employers often receive an enormous number of applications (usually unstructured resumes) that are difficult to process and analyze manually. To address this issue, several automatic recruitment systems have been proposed. Although these systems have proved to be more effective in processing candidate resumes and matching them to their relevant job posts, they still suffer from low precision due to limitations of their underlying techniques. On the one hand, approaches based on keyword matching ignore the semantics of the job post and resume contents, and consequently a large portion of the matching results is irrelevant. On the other hand, the more recent semantics-based models are influenced by the limitations of the used semantic resources, namely the incompleteness of the knowledge captured by such resources and their limited domain coverage. In this paper, we propose an automatic online recruitment system that employs multiple semantic resources to highlight the semantic contents of resumes and job posts. Additionally, it utilizes statistical concept-relatedness measures to further enrich the highlighted contents with relevant concepts that were not initially recognized by the used semantic resources. The proposed system has been instantiated and validated in a precision-recall based empirical framework.

In Near-Duplicate Video Retrieval (NDVR), recent works have focused on bettering index structures and matching schemes not only to improve retrieval accuracy but also to enforce scalability in an effort to keep up with the ever-growing size of video collections. In this paper, we propose a framework for the retrieval of Near-Duplicate Videos (NDVs) based on a pre-processing step of clustering inspired by Multiple Sequence Alignment (MSA) of DNA sequences. In our technique, we represent videos as alphabetical genomes and MSA is employed to automatically cluster a video collection. NDVR is then conducted on these formed clusters instead of the original video collection. Experimentally, we show that our clustering-based approach, while being significantly faster than state-of-the-art techniques that are not based on a pre-processing clustering step, i.e. the n-gram and Dynamic Time Warping (DTW) techniques, yields equivalent results in a precision/recall evaluation framework.

Existing image retrieval frameworks use the low-level visual features extracted from images to learn a set of semantic categories and tend to discard the conceptual properties of these features. In this paper, we propose to use the conceptual information of the low-level image features to establish a correspondence between the image features and a set of high-level semantic categories. In doing so, image objects are indexed using concepts that describe the low-level color, texture, and shape features in a way that reflects the human understanding of their visual properties. Experimentally, we show that using the conceptual information outperforms using the low-level features for indexing and classifying image regions. In the retrieval process, we demonstrate the effectiveness of using the conceptual information in queries to enrich the search process of users by handling non-trivial queries incorporating both the semantic concepts and the conceptual information of the queried semantic concepts.

With the growing development of heterogeneous domain-specific ontologies, the treatment of the semantic and structural differences between such ontologies becomes more important. In addition, constant maintenance and update is required so that they can be promptly enriched with new concepts and instances. In this paper, we present a coupled statistical/semantic framework for ontology merging and enrichment. First, we prioritize the ontology merging techniques according to their significance and execution into semantic-based, name-based, and statistical-based techniques respectively. In addition, we exploit multiple knowledge bases to support the merging task. Second, we use the massive amount of information encoded in texts on the Web as a corpus to enrich the merged ontology. An experimental instantiation of the framework and comparisons with state-of-the-art syntactic and semantic-based merging and enrichment systems validate our proposal.

We propose a framework highlighting symbolic shape concepts based on the characterization of their geometrical properties. Starting from seven basic shapes, we define transformations to generate novel shapes and discuss their organization within a lattice-based structure. These are then automatically assigned a conceptual representation after: (i) the extraction of low-level shape features based on Fourier descriptors, (ii) the mapping of the low-level features with shape concepts through a support vector matching architecture featuring a radial basis function kernel. Experimentally, we compute the accuracy of the symbolic shape characterization through 5-fold cross validation and demonstrate the effectiveness of the shape concepts for symbolic image retrieval. We indeed show, in a recall-precision evaluation framework, that our approach outperforms a state-of-the-art content-based image retrieval architecture based on query-by-example.

Building accurate models for video event classification is an important research issue since they are essential components for effective video indexing and retrieval. Recently kernel-based methods, particularly support vector machines, have become popular in multimedia classification tasks. However, in order to use them effectively, several factors that hinder accurate classification results, such as feature subset selection and selection of the SVM kernel parameters, must be addressed through the use of heuristic-based techniques. We present a new approach to enhance the performance of SVM for video events classification based on a search method. The latter relies on the simultaneous optimization of the feature and instance subset and SVM kernel parameters, with genetic algorithms. Classification results on sport videos show the significant improvement over conventional SVM.

Discovering semantic correspondences between ontology elements is a crucial task for merging heterogeneous ontologies. Most ontology merging tools use several methods to aggregate and combine similarity measures. In addition, some of the ontology merging systems exploit external resources such as, Linguistic Knowledge Bases (e.g. WordNet) to support this task. However, the quality of their results is subjected to the limitations of the exploited knowledge base. In this paper, we present a framework that exploits multiple knowledge bases that cover information in multiple domains for: i) Indentifying and correcting incorrect semantic relations between the concepts of domain-specific ontologies. This is a primary step before ontology merging; ii) Merging domain-specific ontologies; and iii) Handling the issue of missing background knowledge in the exploited knowledge bases by utilizing statistical techniques. An experimental instantiation of the framework and comparisons with state-of-the-art syntactic and semantic-based systems validate our proposal.

In this paper, we propose the use of motion tracks to generate video features, which are applied to perform video retrieval. We simply use the location of the object's center-of-gravity and its velocity across consecutive video frames to from motion flows, which are then recorded and stored in a video database. In the video retrieval phase, we propose the use of n-gram matching strategy to execute the video retrieval task. The retrieval process can be triggered query-by-example. We identify corresponding query and index contents accurately in order to detect "similar" videos. Experiments are carried out on CMU database. An empirical comparison over the state-of-the-art dynamic programming techniques is encouraging and demonstrates the advantage and feasibility of the n-gram method.

In surveillance applications, human gait data obtained from video contains idiosyncratic tendencies which allows it to be used as a biometric. This gait data has both time and image information. Expertise in the domain of time series analysis can be fruitfully employed in the image processing domain. In this paper, we consider the monocular frontal view of gait. In this view we track body parts to obtain time information and in doing so, complete occlusion of body parts may occur. To compensate for this, we present a novel standpoint where occluded images of objects may be considered as data missing from a time series. Thus we can consider this as a new application of the ¿missing data¿ problem studied in other fields dealing with time series data applied to the classic computer vision problem of occlusion. Using this approach, we consider three ways of compensating for occlusion - namely polynomial interpolation, autoregressive prediction and coupled time/frequency domain interpolation. We propose an experimental instantiation using a gait dataset and analyzing the motion of colored markers attached to body parts. The actual and predicted positions are compared which show our approach holds promise for complete occlusion compensation.

Web images come in hand with valuable contextual information. Although this information has long been mined for various uses such as image annotation, clustering of images, inference of image semantic content, etc., insufficient attention has been given to address issues in mining this contextual information. In this paper, we propose a webpage segmentation algorithm targeting the extraction of web images and their contextual information based on their characteristics as they appear on webpages. We conducted a user study to obtain a human-labeled dataset to validate the effectiveness of our method and experiments demonstrated that our method can achieve better results compared to an existing segmentation algorithm.

In this paper, we describe methods for identification of rebroadcasted video sequences. The methods are applicable to real-time monitoring of video streams, or to searching of large collections. The techniques we present are fast and insensitive to video degradation, making them ideal for re-broadcasted video search. Experiments on large quantities of real-world video data demonstrate that our approach is robust and effective, even for short queries or queries that have been edited or degraded. By using dynamic programming and N-gram matching approaches we identify regions of similarity in video signatures efficiently and accurately in order to identify rebroadcasted videos, even when these regions constitute only a small section of the clip being searched.

Gait is an emerging biometric showing promise in its use. Most research focusses on fronto-parallel (FP) gait where people walk across a camera. In this paper, we present an original analysis, presenting the case for the use of fronto-normal (FN) gait where motion is towards a camera. In FN gait, the image projected on a camera sensor will get larger, in a looming effect. This affects the data in a nonlinear and non stationary way, which will further complicate analysis of movement. By catering for this effect, we present new insights into perspective motion compensation for FN gait. Using an existing database which uses coloured markers, we compare two methods of compensation for looming. Initial examination of the resulting data shows a significant result, that fundamentally different approaches may be used for time series analyses on the a set of FN gait data. This opens up new avenues for biometric research in gait recognition.

We propose to enhance the performance of the S.I.R. image indexing and retrieval architecture [1,2] through the integration of a query-by-example (QBE) framework based on high-level image descriptions instead of their extracted low-level features. This framework features a bi-facetted conceptual model coupling visual semantics and relational spatial characterization and operates on image objects (abstractions of visual entities) in an attempt to perform querying operations beyond state-of-the-art relevance feedback (RF) frameworks. Also, it manipulates a rich query language consisting of several boolean operators, which therefore leads to optimized user interaction and increased retrieval performance.

The growing need for "intelligent" image retrieval systems to cope with the fast development of the digital imaging technology leads to new architectures integrating multiple sources of description of the image content. Indeed, addressing the problem of combining modalities for indexing and querying is of huge importance in forthcoming retrieval methodologies and is the only solution for achieving significant retrieval performance on image documents. This paper presents a multi-facetted conceptual framework integrating semantics, texture and spatial features for automatic image retrieval. It features a high-level representation formalism handling symbolic image descriptions and a unified full-text query framework. Experiments with queries combining multiple characterizations of the image content on a corpus of 2500 photographs validate our proposal.

We propose in this paper to enhance the performance of the S.I.R. (Belkhatir et al., 2004; Belkhatir, 2005; Belkhatir et al., 2005) (signal/semantic integration for image retrieval) image indexing and retrieval architecture through the integration of a query-by-example framework based on high-level image descriptions instead of their extracted low-level features. This framework features a multi-facetted conceptual model which integrates visual semantics as well as symbolic color and texture characterizations and operates on image objects (abstractions of visual entities within a physical image) in an attempt to perform querying operations beyond trivial low-level processes and region-based frameworks. Also, it manipulates a rich query language, consisting of both Boolean and quantification operators, which therefore leads to optimized user interaction and increased retrieval performance. Experimental results on a test collection of 2500 color personal photographs show that our approach gives better results in terms of recall and precision measures than state-of-the-art frameworks loosely coupling keyword-based query modules and relevance feedback processes operating on low-level features