Search

We present a novel multi-modal evidence fusion method for highlevel feature (HLF) detection in videos. The uni-modal features, such as color histogram, transcript texts, etc, tend to capture different aspects of HLFs and hence share complementariness and redundancy in modeling the contents of such HLFs. We argue that such inter-relation are key to effective multi-modal fusion. Here, we formulate the fusion as a multi-criteria group decision making task, in which the uni-modal detectors are coordinated for a consensus final detection decision, based on their inter-relations. Specifically, we mine the complementariness and redundancy inter-relation of uni-modal detectors using the Ordered Weighted Average (OWA) operator. The `or-ness΄ measure in OWA models the inter-relation of uni-modal detectors as combination of pure complementariness and pure redundancy. The resulting weights of OWA can then yield a consensus fusion, by optimally leveraging the decisions of uni-modal detectors. The experiments on TRECVID 07 dataset show that the proposed OWA aggregation operator can significantly outperform other fusion methods, by achieving a state-of-art MAP of 0.132. [ABSTRACT FROM AUTHOR]

Feature selection technology can help to reduce feature redundancy and improve classification performance. Most general feature selection methods do not perform well on high-dimension large-scale data sets of multimedia applications. In this paper we propose a novel feature selection method named Local Separability Assessment. We try to measure the separation level of samples in subregions of feature space, and integrate them for evaluating the separability of features. Our method has favorable performance on large-scale continuous data sets, and requires no priori hypothesis on data distribution. The experiments on various applications have proved its excellence. [ABSTRACT FROM AUTHOR]

Text information, which plays important role in news video concept detection, has been ignored in state-of-the-art movie analysis technology. It is so because movie subtitles are speech of roles which do not directly describe content of movie and contributes little to movie analysis. In this paper, we import collaborative-editing synopsis from professional movie sites for movie analysis, which gives detailed descriptive text of movie. Two aligning methods, subtitle alignment and RoleNet alignment, are proposed complementarily to align synopsis to movie to get scene-level text information of movie. The experiment show that proposed methods can effectively align synopsis to movie scene and the imported text information can give a more user-preferred summarization than merely using audiovisual feature. [ABSTRACT FROM AUTHOR]

Image hashing is an alternative approach to many applications accomplished with watermarking. In this paper, we propose a novel image hashing method in the DCT Domain which can be directly extended to MPEG video without DCT transforms. A key goal of the method is to produce randomized hash signatures which are unpredictable for unauthorized users, thereby yielding properties akin to cryptographic MACs. This is achieved by encryption of the block DCT coefficients with chaotic sequences. After applying Principal Components Analysis (PCA) to the encrypted DCT coefficients, we take the quantized eigenvector matrix (8 × 8) and 8 eigenvalues together as the hash signature, the length of which is only 72 bytes for any image of arbitrary size. For image authentication, we also present an algorithm for locating tampering based on the hashing method. Experiments on large-scale database show that the proposed method is efficient, key dependent, pairwise independence, robust against common content-preserving manipulations. [ABSTRACT FROM AUTHOR]

In this paper, we present a novel fingerprinting method for image authentication, the fingerprint length of which is very short (only 81 bytes) and independent on image sizes. First, we extract features based on the block DCT coefficients of images, and binarize the feature map to get key blocks. Then we apply Principal Components Analysis (PCA) to the DCT Coefficients of key blocks. Finally, we take the quantized eigenvector matrix (9×9) as fingerprints. Experimental results show that the proposed method is discriminative, robust against compression, and sensitive to malicious modifications. [ABSTRACT FROM AUTHOR]

With digital sports video increasing everyday, effective analyzing sports video content becomes more and more important. Effective and efficient representation of video for searching, retrieval, inference and mining is a key problem in knowledge engineering. To describe sports video content efficiently, sports video ontology for video annotation is represented in OWL, a description logic based Web Ontology Language. We describe a user-friendly platform for sports video annotation. Ontology based sports video annotation can facilitate video indexing, retrieval and reasoning in a broad range of applications including Digital Olympic Project in China. Moreover, we present a hierarchical sports video summarization strategy to browse the sports video in a progressive way. In sports video, replay scenes often represent the highlight or interesting event of the video. Hence, our representative scene selection is based on the replay detection algorithm and identical events detection. The basic experimental results show our strategy is effective. [ABSTRACT FROM AUTHOR]