Catálogo de publicaciones - libros

The multi-modal characteristics of Web image make it possible to unify keywords and visual features for image retrieval in Web context. Most of the existing methods about the integration of these two features focus on the interactive relevance feedback technique, which needs the user’s interaction (i.e. a two-step interactive search). In this paper, an approach based on association rule and clustering techniques is proposed to unify keywords and visual features in a different manner, which seamlessly implements the integration within one-step search. The proposed approach considers both (QBK) mode and (QBE) mode and need not the user’s interaction. The experiment results show the proposed approach remarkably improve the retrieval performance compared with the pure search only based on keywords or visual features, and achieve a retrieval performance approximate to the two-step interactive search without requiring the user’s additional interaction.

In this paper, a dynamic bandwidth allocation algorithm is proposed for stored video transmission with renegotiation interval constraint. It is to handle the problem of short renegotiation intervals in optimal smoothing algorithms[4,5], which may increase the renegotiation cost or cause renegotiation failures. Based on the transmission rate bounds derived from buffer constraints, a transmission segment is calculated based on the optimal smoothing algorithm [5]. If the length of the segment is less than the minimum renegotiation interval, it is merged to the neighboring segment considering the relation between the transmission rates of neighboring segments by allowing encoder buffer underflows. From the simulation results, the proposed algorithm is shown to keep the renegotiation intervals larger than the minimum and the renegotiation cost is greatly reduced with slight decrease in the channel utilization.

In this paper, we focus on the parameter identification problem, one of the most essential problems in the 3D cloth animation created by multimedia software. We present a novel interactive parameter identification framework which integrates the industry knowledge. The essential of this paper is that we design a hybrid intelligent learning system using statistical analysis of kawabata evaluation system(KES) data from fabric industry database, fuzzy system and radial basis function(RBF) neural networks. By adopting our method the 3D cloth animator can interactively identify the parameters of cloth simulation with subjective linguistic variables while in the past decades it is very difficult for cloth animators to tune the parameters. We solve the 3D cloth parameter problem using the intelligent knowledge integration method for the first time in the multimedia and graphics research area and our method is applied to the most popular 3D tool Maya. The experimental results illustrate the practicability and expansibility of this method.

Multi-view video is becoming increasingly popular, as it provides users greatly enhanced viewing experience. Multi-view video coding (MVC) focuses on exploiting not only the temporal correlation among the adjacent pictures for each view, but also inter-view correlation. Though the coding efficiency is a key target for MVC, the view-temporal prediction structure for improving the compression efficiency usually results in the decoding delay and limits the random access ability. Random access ability is an important feature in MVC because it provides the view switching, temporal frame sweepingly browsing and other interactive abilities for the client users in multi-view video streaming. In this paper, we propose an algorithm to flexibly regulate the viewtemporal prediction structure. It is able to achieve a good trade-off between compression performance and random access ability.

This paper presents a novel solution for efficient representation of multi-channel spatial audio signals. Unlike other spatial audio coding techniques, the solution inherently requires no additional side information to recover the surround sound panorama from a two-channel downmix. For a typical five-channel case, only a stereo downmix signal is required for the decoder to reconstruct the full five-channel audio signal. In addition to the bandwidth saved by transmitting no side information, the technique has significant advantages in terms of computational complexity.

A new approach to analyze and synthesize texture regions in video coding is presented, where texture blocks in video sequence are synthesized using graph cut technique. It first identifies the texture regions by video segmentation technique, and then calculates their motion vectors by motion vector (MV) scaling technique like temporal direct mode. After the correction of these MVs, texture regions are predicted from forward and/or backward reference frames by the corrected MVs. Furthermore, Overlapped Block Motion Compensation (OBMC) is applied to these texture regions to reduce block artifacts. Finally, the texture blocks are stitched together along optimal seams to reconstruct the current texture block using graph cuts. Experimental results show that the proposed method can achieve compared visual quality for texture regions with H.264/AVC, while spending fewer bits.

Real-time video transmission over packet networks faces several challenges such as limited bandwidth and packet loss. Multiple description coding (MDC) is an efficient error-resilient tool to combat the problem of packet loss. The main problem of MDC is the mismatch of reference frames in encoder and decoder, when some descriptions are lost during transmission. This paper presents an adaptive error recovery (AER) scheme for multiple description video coding. The proposed AER scheme, which is based on statistical analysis, can adaptively determine the nearly optimal error recovery (ER) method among our predefined ER methods such as interpolation, block replacement and motion vector (MV) reusing. The AER scheme has three advantages. First, it efficiently reduces the mismatch error. Second, it is completely based on pre- post-processing which requires no modification of the source coder. Third, it has low computational complexity, which is suitable for real-time video applications. Simulation results demonstrate that our proposed AER scheme achieves better performance compared with MDC with fixed error recovery (FER) scheme over lossy networks.

The motion vector prediction (MVP) is an important part of video coding. In the original median predictor, if the neighbor blocks of current block are intra-mode coded, their motion vectors (MVs) will be set to zeros for MVP of current block. This is not very precise for sequences with strong motion. This paper propose an improved motion vector prediction (MVP) scheme for H.264. In the proposed scheme, when there are intra-mode macroblocks beside current block, more MV of the neighbor inter-mode block is utilized instead of zero MVs of intra-mode macroblocks for MVP of current block. The experimental results show that the improved scheme achieves better coding efficiency than the original median predictor. Meanwhile the point obtained by the proposed MVP scheme is closer to the global minimum point, the following fast motion estimation (FME) computation complexity is reduced.

This paper proposes an effective time-series classification model based on the Neural Networks. Classification under this model consists of three phases, namely , , and . The main contributions of the paper are described as following: We propose a feature extraction algorithm, which involves computation of finite difference of sequences, for preprocessing. We employ two different types of Neural Networks for training and testing. The results of the experiments on real univariate motion capture data and synthetic data show that our approach is effective in providing good performance in terms of accuracy. It is therefore a promising method for classifying time-series, in particular for univariate human motion capture data.

The demand for information services considering personal preferences is increasing. In this paper, aiming at the development of a system for automatically acquiring personal preferences from TV viewers’ behaviors, we propose a method for automatically estimating TV viewers’ intervals of interest based on temporal patterns in facial changes with Hidden Markov Models. Experimental results have shown that the proposed method was able to correctly estimate intervals of interest with a precision rate of 86.6% and a recall rate of 80.6%.