Catálogo de publicaciones - libros

Many researchers try to utilize the semantic information extracted from visual feature to directly realize the semantic video retrieval or to supplement the automated speech recognition (ASR) text retrieval. But bridging the gap between the low-level visual feature and semantic content is still a challenging task. In this paper, we study how to effectively use Latent Semantic Indexing (LSI) to improve the semantic video retrieval through the ASR texts. The basic LSI method has been shown effective in the traditional text retrieval and the noisy ASR text retrieval. In this paper, we further use the lexicon-guided semantic clustering to effectively remove the noise introduced by news video’s additional contents, and use the cluster-based LSI to automatically mine the semantic structure underlying the terms expression. Tests on the TRECVID 2005 dataset show that the above two enhancements achieve 21.3% and 6.9% improvements in performance over the traditional vector-space model(VSM) and the basic LSI separately.

We propose an adaptive method that can estimate 3D position of a soccer ball by using two viewpoint videos. The 3D position of a ball is essential to realize a 3D free viewpoint browsing system and to analyze of soccer games. At an image processing step, our method detects the ball by selecting the best algorithm based on the ball states so as to minimize the chance to miss the ball and to reduce the computation cost. The 3D position of the ball is then estimated by the estimated 2D positions of the two camera images. When it is impossible to obtain the 3D position due to the loss of the ball in an image, we utilize the Kalman Filter to compensate the missing position information and predict the 3D ball position. We implemented a preliminary system and succeeded in tracking the ball in 3D at almost on-line speed.

Particle filtering provides a general framework for propagating probability density functions in non-linear and non-Gaussian systems. However, generic particle filter (GPF) is based on Monte Carlo approach and sampling is a problematic issue. This paper introduces a parzen particle filter (PPF) which uses a general kernel approach to better approximate the posterior distribution rather than Dirac delta kernel in GPF. Furthermore, we adopt multiple cues and combine texture described by directional energy from multi-scale, multi-orientation steerable filtering with color to characterize our tracking targets. The advantages of tracking with multiple cues compared to individual ones are demonstrated over experiments on artificial and natural sequences.

A novel method for creating diverse ensembles of image classifiers is proposed. The idea is that, for each base image classifier in the ensemble, a random image transformation is generated and applied to all of the images in the labeled training set. The base classifiers are then learned using features extracted from these randomly transformed versions of the training data, and the result is a highly diverse ensemble of image classifiers. This approach is evaluated on a benchmark pedestrian detection dataset and shown to be effective.

In this paper, we propose a hybrid content-based image authentication scheme that integrates two complementary algorithms: Robust content-based authentication and semi-fragile crypto-hash based authentication. The former uses global features and is quite robust against various types of noise. The latter uses local features and therefore is able to identify the tempered area in case the image is attacked. The proposed scheme takes advantage from both algorithms and provides more information to guide the decision maker. In addition, we also propose two improved algorithms based on Fridrich’s content-based and Sun’s crypto-hash based authentication. Experiments show that the improved algorithms are more secure than the original algorithms. Another contribution of this paper is that, by concatenating the signatures generated with two different authentication algorithms, the fuzzy area in authentication decision can be further quantized, which provides more choices for authentication decision.

P2P networks play a promotive role in distribution and transmission of digital multimedia content by providing high availability, fault tolerance, bandwidth efficiency and dynamic scalability. At the same time, however, it facilitates illegal pirate and unauthorized access towards copyright media content which may violate possessor ownership and result in economic loss. Conventional client/server model for DRM has difficulties coping with the challenge of secure communication in P2P environment. In this paper, we propose a strategy to incorporate DRM mechanism into P2P network architecture and construct a system which ensures efficient content sharing as well as reliable media protection. Our approach implements a revocable key management scheme based on broadcast encryption and enables cryptographic information for session key processing to be distributed among users within the same authorized domain in peer-to-peer mode. Mathematical analysis has demonstrated that the new strategy outperforms traditional solutions on alleviating communication overhead of License Server, minishing peer latency of authorization, and importing security modules without much modification to original peer-to-peer infrastructure.

With the increase of video applications, the security of video data becomes more and more important. In this paper, we propose a new video encryption scheme for H.264/AVC video coding standard. We define Unequal Secure Encryption (USE) as an approach that applies different cryptographic algorithms (with different security strength) to different partitions of video data. The USE scheme includes two parts: video data classification and unequal secure video data encryption. For data classification, we propose 3 data classification methods and define 5 security levels in our scheme. For encryption, we propose a new stream cipher algorithm FLEX and XOR method to reduce computational cost. In this way, our scheme can achieve both high security and low computational cost. The experimental results show that the computational cost of the USE scheme is very low. In security level 0, the computational cost is about 18% of naive encryption. The USE scheme is very suitable for high security and low cost video encryption systems.

Tattoos are used by law enforcement agencies for identification of a victim or a suspect using a false identity. Current method for matching tattoos is based on human-assigned class labels that is time consuming, subjective and has limited performance. It is desirable to build a content-based image retrieval (CBIR) system for automatic matching and retrieval of tattoos. We examine several key design issues related to building a prototype CBIR system for tattoo image database. Our system computes the similarity between the query and stored tattoos based on image content to retrieve the most similar tattoos. The performance of the system is evaluated on a database of 2,157 tattoos representing 20 different classes. Effects of segmentation errors, image transformations (e.g., blurring, illumination), influence of semantic labels and relevance feedback are also studied.

Knowledge of paintings domain includes a variety of sources such as essays, visual examples, ontologies of artistic concepts and user- provided annotations. This knowledge serves several purposes. First, it defines a wide range of concepts for annotation and flexible retrieval of paintings. Second, it serves to bootstrap auto-annotation and disambiguate the generated candidate labels. Third, the user-provided annotations serve to discover folksonomies of concepts and vernacular terms. In this paper, we propose a framework for paintings auto-annotation that incorporates user provided images and annotations, domain ontology and external knowledge sources. We utilize these sources of information to bootstrap and support the auto-annotation task, which is based on transductive inference mechanism that combines probabilistic clustering and multi-expert approach to generate labels. We further combine user-provided annotations with generated labels and domain ontology to disambiguate the concepts. In our experiments, we focus on the auto-annotation of painting and demonstrate that the user-provided annotations significantly increase annotation accuracy.

Value Combination Technique is proposed for a novel bitmap exact authentication. It combines the pixel values in an image block with the block position, and the combined values are used as the initial state of chaotic system to generate watermark. Furthermore, a general rule is given to analyze the reliability of algorithm, and another rule is presented to design an algorithm with high reliability in this paper. Then, a concrete algorithm illustrates our proposed authentication system. Extensive experiments show that this system can effectively resist such as feature extraction attack, vector quantization attack and so on, and be very sensitive to tamper.