Catálogo de publicaciones - libros

Tone analysis is the first step in investigations of Chinese dialects. Traditionally, this is done by well-trained linguists. In this paper, a SOM-based automatic tone analysis approach is proposed. SOM is facilitated to map the F0 contour to 2-dimensional space, and then the data are clustered to indicate tones of the dialect. We also introduce a phonology factor to the DB-index, which is used to validate the clustering, in order to incorporate phonology knowledge. Our version of DB-Index assumes that characters that have similar tones in the past tend to have similar tones today. Experiments show that the approach attains satisfactory results.

Now, techniques for content-based image retrieval are not yet mature enough to recognize visual semantics completely. To interpret semantic of image, many researchers use keywords as textual annotation. However, it’s the image retrieval without ranking by text matching which is the simplest way to retrieval according to keyword’s existence or nonexistence. In other words, dissimilarity computation among terms is usually done by using string matching not concept matching. In this paper, we change a simple keyword set to conceptualization through similarity measure between keywords in the annotated images. After conceptualization, we extract major and minor keyword and apply them to concept-based image retrieval and classification. In the future work, we can add new keywords(missing keywords because of annotator’s different point of view) or re-arrange keywords.

In this paper, a retrieval system which enables to retrieval digital video and image compressed by MPEG and database divided 6 categories is implemented. Video retrieval detects content-based features value by using differential value of pixels, chi-square values on histogram images, differential value of histogram’s variance and chi-square value of variance about each bin which is histogram of quantized image. After detecting the shot boundary by the feature value of mean value comparison algorithm, content-based feature value is detected from key frame representing each shot by using HMMD color space, Edge Histogram and category image retrieval detect feature value by applying MPEC-7 standard visual descriptors. The detected feature values are neural network in the engineering domain, and learnt content-based feature vectors fast and apply the hippocampal neural network algorithm to compose of optimized feature. We present fast and precise retrieval effect when indexing and retrieving.

This paper will be addressing why communication between M2M(machine to machine) in USN(Ubiquitous Sensor Network) has become a hot issue in IT recently. We propose RFID(Radio Frequency Identification) biometrics systems for personal certification to communicate between M2M. This system stores extracted feature vectors from a facial image to each tag. It performs a comparison and interpretation when facial image is inputted. If bio-information of the tag is in accordance with the extracted feature vectors from the input facial image, it is possible to access the database. Otherwise, authentication has failed. The proposed system is expected to be available for personal authentication in the USN with combination of biometrics, which is face, fingerprint, iris, etc, and RFID system.

Mosaicing of microscopic images is often necessary when the observed specimen cannot be captured into a single image. Automatic method is preferred because it will greatly reduce the work involved. In the paper, we present a feature based automatic mosaicing method based on the related research on panorama reconstruction for photography. Scale invariant feature transform (SIFT) is first applied to extract robust features from the images, and by careful implementation of Best-Bin-First (BBF) algorithm, we construct the global k d-Tree from all the features and search for the possible overlapping image pairs efficiently. Random sample consensus (RANSAC) is chosen to further verify the matches. And once the image pairs are all validated, minimum spanning tree (MST) is used to obtain the best connected-component of the image set to recover the transformation between images and project them into the mosaic frame. Our experiment results show that the approach is robust to background noises and illumination change in the images and can give reliable and accurate results even for images of low overlapping or with relatively few features.

In the video coding process, motion estimation is one of the most computational problem because of enormous amount for calculation. Especially, the recent MPEG-4 AVC (advanced video coding) standard requires much more computations in motion estimation than the conventional MPEG-2 coding standard. In this paper, we propose a fast partial distortion elimination algo-rithm using adaptive matching scan and refining threshold which can reduce only unnecessary computations significantly. Our algorithm is based on selec-tive matching scan and elimination of unlike candidate blocks from initial matching error and the relation of SAD(Sum of Absolute Difference) between the current block and correlated blocks. According to the obtained matching order, we proceed to calculate matching errors and remove unlike candidate vectors using adaptive threshold based on PDE (Partial Distortion Elimination) method. The adaptive threshold is obtained by relations already computed dis-tortion statistics. Our algorithm takes about 4~10% of computations for block matching error compared with conventional FS algorithm, thus our algorithm will be useful to real-time video coding applications using MPEG-4 AVC or MPEG-2 video coding standards.

In this paper, a new steganalysis approach is proposed to reliably detect hidden message in image and estimate the amount of it. The scheme is based on the fact that LSB embedding can increase the number of close pairs of colors, and the statistical analysis of information entropy of image. Thus the change of entropy value in image can reflect the being of hidden message. Experiments show that the approach leads to good estimates of the hiding ratio.

A scheme of noise canceler corresponding to nonlinear path using a parallel adaptive filter (PAVF) is presented in this paper. The adaptive Volterra filter (AVF) corresponding to nonlinear systems is attractive due to its linear relationship between the input and output signals. However, its formidable computer complexity is prohibitive for practical applications. In this paper, instead of the usual AVF, a design method of noise canceller using PAVF is presented. PAVF consists of several subfilters which is partitioned from AVF and can reduce the computational work significantly. Simulation results validated our proposed method.

This paper proposes a recursive parametric spectral subtraction (RPSS) algorithm based on the characteristics of the human auditory system to calculate the masking threshold. Since the masking threshold of a clean signal is difficult, if not impossible, to calculate from noisy speech, we add a recursive process to estimate the subtraction factor and the spectral floor factor of parametric spectral subtraction approaches. The recursive process is terminated when the output SNR of the current signal frame shows no further improvement. The effectiveness of the proposed RPSS has been validated by the SNR improvement test and the recognition rate test and compared with the standard power spectral subtraction (PSS) method. According to the experimental results, the proposed method outperforms the PSS method in both tests.

Many algorithms based on wavelets have been shown to work well in image compression. However, the results are not good enough to meet the applications because wavelets cannot simultaneously possess all of the desirable properties such as symmetry, orthogonality, compact support and high approximation order, etc. Offering these features simultaneously as the extension from wavelets, multiwavelets can give better results in signal denoising as well as image compression. This paper will show a method of a quantizer for image compression of algae cell using multiwavelets, which simply combines the SPIHT and EZW coder. The experimental results demonstrate that our method exhibits better performance than the existing ways for image compression such as scalar wavelets or other traditional transforms.