Showing papers presented at "IEEE-Eurasip Nonlinear Signal and Image Processing in 2005"

New method for noise estimation in images

Abstract: Summary form only given. In this paper, we propose a simple and fast algorithm for additive noise estimation. In our approach, the input image is assumed to be corrupted by white additive Gaussian distributed noise. To estimate the noise variance, a three step method is proposed that posses good performances at a low computational cost. Simulations results that show the performance of the proposed implementation in comparison to other approaches are also presented in this paper.

A multichannel weighted median filter for complex array signal processing

Abstract: Summary form only given. Multi-channel and multi-spectral signals are often correlated across channels. Moreover, multispectral images have considerable similarity in in-channel correlations. But in array signal processing, due to the existence of multiple frequency components and their phase shifts, in-channel correlations may vary drastically. In this paper, a new multichannel weighted median filter is proposed which can capture the general correlation structure in array signals and process them in an efficient manner. The algorithm is further extended onto the complex domain by means of "phase coupling". The performance of the filter is presented in a three-sensor array processing example.

Two-stage methods for mixed noise removal [image denoising]

Abstract: Summary form only given. Images are quite often corrupted by mixed additive and impulsive noise. Then, the task in the filtering is to remove both components of the noise. We compare in this paper two ways how to do this. The first one is to apply some robust order statistic filter and the other one is to use a two-stage approach where impulses are first detected and removed, and after that additive noise is suppressed. For the latter approach, two methods are proposed. We demonstrate through experiments that the latter approach performs better than several standard order statistic filtering techniques. In addition, a modified version of a method to estimate the variance of additive noise is introduced. The given method performs well enough even with mixed noise.

Application of a double-talk resilient DFT domain adaptive filter for bin-wise stepsize controls to adaptive beamforming

Abstract: Summary form only given. In adaptive filtering, undetected noise bursts often disturb the adaptation and may lead to instabilities and divergence of the adaptive filter. The sensitivity against noise bursts increases with the convergence speed of the adaptive filter and limits the performance of signal processing methods where fast convergence is required. Typical applications which are sensitive against noise bursts are adaptive beamforming for audio signal acquisition or acoustic echo cancellation, where noise bursts are frequent due to undetected double-talk. In this paper, we apply double-talk resistant adaptive filtering (Gaensler (1998)) using a nonlinear optimization criterion to adaptive beamforming in the discrete Fourier transform domain for bin-wise adaptation controls. We show the efficiency of double-talk resilient adaptive filtering for a generalized sidelobe canceller for speech and audio signal acquisition. The improved robustness leads to faster convergence, to higher noise reduction, and to a better output signal quality in turn.

A AM/FM single component signal reconstruction using a nonsequential time segmentation and polynomial modeling

Abstract: Summary form only given. The problem of estimating nonstationary signals has been considered in many previous publications. In this paper we propose an alternative algorithm in order to accurately estimate AM/FM signals. Only single component signals are considered. We perform local polynomial modeling on short time segments using a nonsequential strategy. The degree of polynomial approximation is limited due to the shortness of each time segment. The time support of a segment is controlled by a criterion defined on the spectrogram. To keep optimality a maximum likelihood procedure estimates the local model parameters leading to a nonlinear equation system in R7. This is solved by a simulated annealing technique. Finally, the local polynomial models are merged to reconstruct the entire signal model. The proposed algorithm enables highly nonlinear AM/FM estimation and shows robustness even when signal to noise ratio (SNR) is low. The appropriate Cramer Rao bounds (CRB) are presented for both polynomial phase and amplitude signals. Monte Carlo simulations show that the proposed algorithm performs well. Finally, our proposed method is illustrated using both numerical simulations and a real signal of whale sound.

Multi-layer neural network for precursor signal detection in electromagnetic wave observation applied to great earthquake prediction

Abstract: Summary form only given. It is well known that the electromagnetic (EM) waves that radiate from the Earth's crust are useful for predicting earthquakes. We observe electromagnetic waves in the extremely low frequency (ELF) band of 223 Hz. These observed signals contain several undesired signals due to fluctuations in the magnetosphere or the ionized layer, lightning radiation from the tropics, and so on. This paper proposes a multilayer neural network (NN) using compression data for precursor signal detection. Input data are reduced by wavelet transform. Moreover, we discuss an implementation of the hidden layer. It is shown that the proposed neural network is useful for precursor signal detection.

GA-based adaptive image watermarking with JND profile and fuzzy inference system

Abstract: Summary form only given. This paper presents a GA-based adaptive image watermarking technique with just-noticeable distortion (JND) profile and fuzzy inference system (FIS) which is referred to as the GAIWJF technique. During watermark embedding, the GAIWJF technique embeds a watermark into an image by referring to the JND profile of an image so that the technique makes the watermark more imperceptible. The GAIWJF technique employs image features and local statistics to create an FIS containing three fuzzy input variables, eight fuzzy inference rules, and a single fuzzy output variable. During watermark extraction, the GAIWJF technique does not require the information of original images because it employs the FIS to extract watermarks. Also, the FIS is further optimized by a GA so that the performance of watermark extraction can be improved. From experimental results, the GAIWJF technique not only makes the embedded watermark more imperceptible but also possesses adaptive and robust capabilities to resist image manipulations.

Marketing data collection from face images using neural networks

Abstract: Summary form only given. In the marketing field, it is very useful to collect the customer data of age and gender etc. For gender and age classification, this study focuses on the edge information that consists of all wrinkles in a face and also a neck. A density histogram in which the value of the edge intensity of the vertical direction and the horizontal direction in an image is greater than a threshold value is computed by using the edge information. They are treated as input data to a neural network. This classification is used on a face image database in which are the collected face images of people 15-64 years old. In order to show the effectiveness of the proposed method, computer simulations are carried out using these face images.

Noise robust speech recognition based on spatial subtraction array

Abstract: Summary form only given. We propose a spatial subtraction array (SSA) and known noise superimposition to achieve a robust hands-free speech recognition under noisy environments. In the proposed SSA, noise reduction is achieved by subtracting the estimated noise power spectrum from the target speech power spectrum to be enhanced in the mel-scale filter bank domain. This offers a realization of error-robust spatial spectral subtraction with few computational complexities. In addition, we introduce known noise superimposition techniques in the mel-scale filter bank domain, and utilize the matched acoustic model for the known noise. This can compensate the acoustic model mismatch and mask the residual noise component in SSA. The experimental results obtained under a real environment reveal that word accuracy of the proposed method is greater than that of the conventional method even when the target user moves between -10 and +10 degrees around the microphone array.

Elastic graph matching versus linear subspace methods for frontal face verification

Abstract: Summary form only given. In this paper, a comparative study between standard linear subspace techniques such as eigenfaces and fisherfaces and a novel morphological elastic graph matching for frontal face verification is presented. A set of experiments has been conducted in the M2VTS database in order to investigate the performance of each algorithm in different image alignment conditions. The experimental results indicate the superiority of the novel morphological elastic graph matching against all the other presented techniques.

Acoustic space analysis method utilizing statistical multidimensional scaling technique

Abstract: Summary form only given In order to achieve sufficient improvement in speaker-adaptation techniques, such as the MLLR method, it is essential to obtain an adequate number of samples of the user's voice, rendering the application of the method difficult in practical environments Prior development of a library of highly precise acoustic models is necessary to ensure high enough speech recognition performance from the outset of using the system It is quite important to analyze a target acoustic space to design an efficient acoustic model library However, the analysis of multidimensional acoustic space is generally a difficult task In order to support the analysis of acoustic space through the capability of human visual perception, we proposed the COSMOS (COmprehensive Space Map of Objective Signal, previously aCOustic Space Map Of Sound) method It features the visualization of an aggregate of acoustic models based on stochastic models, such as HMM and GMM, into a two-dimensional map (called COSMOS map) by utilizing a statistical multidimensional scaling technique of nonlinear projection First, the paper formulates the COSMOS method Then, a quantitative analysis of a speaking style COSMOS map is described Error analysis of the mapping from multidimensional space to two-dimensional space in the COSMOS map is investigated Furthermore, it is suggested that there exist multiple radiated axes of acoustic feature continuity in the COSMOS map

Colored digital watermarking technology based on visual cryptography

Abstract: Summary form only given. For a period of time, it was popular to use gray-level and black-and-white authentication logos for digital watermarking. With the changes of times, there are more and more digital watermarking techniques using colored logos for the index of its authentication. However, the algorithm must be more complicated for using colored logos to develop digital watermarking. It depends on a lot of complicated mathematical operations to embed the colored digital watermark. In this paper, an application of using visual cryptography technology on the making of colored digital watermarks is proposed. The proposed scheme is promoted by the great advantage of visual cryptography as it is without complicated mathematical operations. Using the human visual system, the secret can be distinguished. Our approaches not only keep the advantage of visual cryptography, but also will not destroy the protected image while making the colored digital watermark.

Subjective and objective quality assessment of noise reduced speech signals

Abstract: Summary form only given Recently, hands-free speech communication is becoming increasingly necessary for teleconferences, in-car telephones, and PC-based IP telephony Since speech signals acquired by a distant microphone are generally corrupted by ambient noise, noise reduction is indispensable for ensuring speech quality To investigate the effect of noise reduction on speech quality, the subjective quality assessment of noisy speech signals and noise reduced speech signals was performed This result confirmed that the difference of noise types gives little effect on the subjective quality and the subjective quality is more sensitive to the speech distortion than the residual noise This paper also investigated the applicability of the PESQ to the objective quality assessment of noise reduction algorithms The experimental result showed that the objective quality correlates well with the subjective quality However, there was the case that the objective quality is lower than the subjective quality in the specific noise reduction algorithms

On the empirical mode decomposition in time-frequency representation of audio signals considering disjoint orthogonality

Abstract: Summary form only given Empirical mode decomposition (EMD) is an adaptive process to decompose a non-stationary and nonlinear signal into the oscillating components The EMD together with Hilbert transformation produces the Hilbert spectrum (HS), a fine-resolution time-frequency representation of non-stationary signals Time-frequency representation is used in many signal-processing techniques, including separation of mixed audio signals In this paper, we have presented the disjoint orthogonal property of audio signals as the decisive factor to measure the efficiency of the time-frequency representation Some audio sources are considered as disjoint orthogonal if not more than one source is active at any time-frequency point The audio signals are projected onto the time-frequency plane using HS and short-time Fourier spectrum and the quantitative measures of disjoint orthogonality with both of the methods are compared The experimental results show that the HS based method performs better in time-frequency representation of the audio signals with the consideration of disjoint orthogonality

Fuzzy regression by asymmetric support vector machines

Abstract: Summary form only given. The paper presents a modified framework of support vector machines, called asymmetric support vector machines (ASVMs), which is designed to evaluate the functional relationship for fuzzy linear and nonlinear regression models. In earlier works, to cope with different types of input-output patterns, strong assumptions were made regarding linear fuzzy regression models with symmetric and asymmetric triangular fuzzy coefficients. However, the nonlinear fuzzy regression model has received relatively little attention, with such models having certain limitations. This study modifies the framework of support vector machines in order to overcome these limitations. The principle of ASVMs is to join an orthogonal vector to a weight vector in order to rotate the support hyperplanes. The supreme merits of the proposed model are its simplicity, understandability and effectiveness. Consequently, experimental results and comparisons are given to demonstrate that the basic idea underlying ASVMs can be effectively used for parameter estimation.

A new method to remove the Gaussian noise from image in wavelet domain

Abstract: Summary form only given. In experiments, the observed image is often modeled as a noisy image. If the image is embedded in an additive Gaussian noise, the classical solution to the noise removal problem is to use the Wiener filter or median filter. In recent years, the BayesShrinkWavelet method has received attention. In this paper, we present a method to remove the noise from an image including substantial Gaussian noise using scaling coefficients and DACWMF (directional adaptive center weighted median filter) procedure which is based on the BayesShrink method. In this way, we can filter the large-amplitude noise which cannot be removed using BayesShrink and improve the quality of the "cleaned" image.

An effective video matching algorithm for adaptive video fingerprints

Abstract: Summary form only given. The proliferation of video content makes video similarity detection an indispensable tool in video management, searching, and navigation. We propose a new algorithm that targets detecting given video clips from TV programs. The algorithm has the following two advantages: it is simple and can be implemented in realtime; it is robust against TV channel noise and the video re-editing frequently seen on TV programs, for example, slow motion. Our algorithm improved video fingerprints, a region based video feature, with an adaptive computation. A novel dynamic time warping algorithm is used to cope with noise and video re-editing. Experimental results demonstrate the effectiveness of the algorithm.

Robust speech recognition under noisy environments based on selection of multiple noise suppression methods

Abstract: Summary form only given. To achieve high recognition performance for a wide variety of noise and for a wide range of signal-to-noise ratio, this paper presents the integration of four noise reduction algorithms: spectral subtraction with smoothing of time direction; temporal domain SVD-based speech enhancement; GMM-based speech estimation; and KLT-based comb-filtering. In this paper, we investigated the optimal suppression method for each noise condition, and then also developed the method of choosing the optimal method automatically for unknown noise. Recognition results on the AURORA-2J task show the effectiveness of our proposed method.

The target tracking using the spatial-temporal probability model

Abstract: Summary form only given. Moving object extraction and tracking are the preliminary and fundamental processes in developing intelligent human-machine interaction and surveillance systems. In conventional target tracking systems, methods of background subtraction are applied to extract moving objects. However, a noisy image may be generated under a non-stationary background. Pixel-based temporal probability models are then proposed to reduce the noise effect, but the misalignment problem during target tracking on mosaic images makes the object extraction process inaccurate. We improve the method of using pixel-based temporal probability models by constructing spatial-temporal probability models to overcome the misalignment problem. Furthermore, the mosaic images are formed by stitching the images captured by an active camera with pan-tilt movements such that the proposed system can extract and track a moving object over a wide area in real-time.

Chrominance signal interpolation of YUV4:2:0 format color images

Abstract: Summary form only given. This paper presents a chrominance signal interpolation method for YUV4:2:0 format color images. The conventional interpolation methods are linear interpolation, such as bilinear interpolation, bicubic interpolation and so on. In the YUV4:2:0 format, the chrominance signal has lost its high-frequency components. The lost high-frequency components cannot be recovered by the linear interpolation technique. The proposed method aims at the recovery of high-frequency components of the chrominance signals using the luminance signal (i.e., Y signal) as the extension of bilinear interpolation. Since, in the case of the YUV4:2:0 format, the luminance signal has high-frequency components. Simulation results show the superior performance of the proposed approach, compared with bilinear interpolation.

Tracking of body status using extended Kalman filter

Abstract: Summary form only given. Detection of human physical status is an important aspect of context awareness of modern health monitoring systems. In this paper, an approach for tracking body status by array signals from wearable body accelerometer sensors is presented. The method presented here is based on extended Kalman filter (EKF). The EKF takes the outputs of activity (standing, lying down and sitting) classifiers as the a priori knowledge of body postures and further computes the precise body posture and movement. This approach has been applied to activity monitoring and the experimental results indicate that the EKF are able to track various human activities with good accuracy.

Estimating embedding parameters using structural learning of neural network

Abstract: Summary form only given The paper presents a novel approach to estimating the embedding parameters for the reconstruction of the underlying dynamic system from an observed nonlinear time series The estimation is performed by a feedforward neural network with variance suppressive learning, a kind of structural learning proposed by the authors earlier It has been found that the proposed method is more efficient than conventional methods for estimating the embedding parameters for reconstruction of the attractor in the phase space The efficiency of the proposed method has also been verified for short term prediction of a nonlinear time series The simulation results show that the neural network predictor with selection of parameters from the knowledge of embedding parameters from the proposed scheme is more stable and needs faster training than the neural network predictor with parameters from conventional methods

An improved correlation-based watermarking method for images using a nonlinear programming algorithm

Abstract: Summary form only given. This paper aims to improve the accuracy of watermark detection in correlation-based watermarking methods for images. In the correlation-based watermarking method, watermark detection is carried out by computing the correlation between a watermarked image and a pseudorandom code to be checked for its presence. Hence its accuracy depends on the characteristics of host images. In this paper, we first show an experimental result that a certain image may cause watermark detection errors. Then we develop an improved version of the watermark embedding process to be effective in such a situation. The improvement is accomplished by designing a whitening filter using a nonlinear programming algorithm. Numerical experiments show good performance as expected by us.

Improvement of the image enlargement method based on the Laplacian pyramid representation

Abstract: Summary form only given. It is necessary to predict unknown higher-frequency components that are lost by sampling for enlarging digital images. Based on the Laplacian pyramid representation, the prediction of unknown higher-frequency components is equivalent to the prediction of a known higher-resolution Laplacian image. We have proposed the higher resolution method based on the Laplacian pyramid representation. However, the Laplacian pyramid representation was considered for image compression. Thus, we think that the bandwidth of the Gaussian filter for image compression is not optimal for digital image enlargement. In this paper, we proposed a new enlargement method for digital images with the variable bandwidth of Gaussian filter.

Interpolated distanced bigram language models for robust word clustering

Abstract: Summary form only given. Two methods for interpolating the distanced bigram language model are examined which take into account pairs of words that appear at varying distances within a context. The language models under study yield a lower perplexity than the baseline bigram model. A word clustering algorithm based on mutual information with robust estimates of the mean vector and the covariance matrix is employed in the proposed interpolated language model. The word clusters obtained by using the aforementioned language model are proved more meaningful than the word clusters derived using the baseline bigram.

Noise cancellation method for OFDM systems

Abstract: Summary form only given. The orthogonal frequency division multiplexing (OFDM) signal is corrupted by a multipath channel and additive white Gaussian noise (AWGN). The OFDM system introduces a guard interval to avoid inter-symbol interference (ISI). However, an equalizer is required to compensate for fading distortion even if the delay of the multipath channel is shorter than the length of guard interval. The frequency domain equalizer is often used in OFDM systems. However, the estimation accuracy of channel response is degraded due to the AWGN. In order to solve the problem, a noise cancelling technique based on an adaptive digital filter is proposed. The OFDM signal is periodic with a period of fast Fourier transform (FFT). So, the current OFDM signal has high correlation with the signal shifted by a period of FFT. On the other hand, the current AWGN is not correlated with the AWGN shifted by a period of FFT. Therefore, the proposed system estimates only the OFDM signal, which is corrupted by the multipath channel, by taking advantage of the difference between autocorrelation of the OFDM signal and the AWGN.

Adaptive type-2 fuzzy median filters for removal of impulsive noises [image denoising]

Abstract: Summary form only given, as follows. Image detail preservation and impulse noise attenuation are not easy to simultaneously achieve in image restoration design. This study proposes a novel adaptive type-2 fuzzy median (type-2 FM) filter to accomplish these objects. The novel filter is mainly based on the uncertainty handling ability of the type-2 fuzzy sets to use the filtering behavior merits of two filters: the fuzzy median (FM) filter and the type-1 FM filter. On the design of a type-2 FM filter, the proposed method adopts a powerful scheme to extend the restriction on the derivation of membership functions by the FM filter, and a flexibility inference mechanism to improve the performance in the limited memory usage by the type-1 FM filter. Extensive simulation results illustrate that the type-2 FM filter not only possesses desirable robustness in suppressing noises but also outperforms other proposed filtering approaches.

Analysis of a nonlinear system via internal-state of a neural network

Abstract: Summary form only given. The internal state of a network has been inspected to evaluate the performance of the network. In particular, the weight vectors of the network have been applied for the analysis of a time series such as biological signals and nonstationary signals. The complexity (eg. nonlinearity and nonstationarity) of such signals often makes it a challenging task to use them in the signal processing field. In this paper, we propose a new neural network based technique to address these problems. We show that a feed forward, multi-layered neural network can conveniently capture the parameter change of a nonlinear system in its connection weight-space, after a process of supervised training. The performance of the proposed method is investigated with a linear and nonlinear system simulated via a mathematical equation.

Improvement of preprocessing method on fingerprint identification system by layered neural networks

Abstract: Summary form only given, as follows. We constructed an individual identification system of fingerprints by three-layered neural networks, and investigated the effects of the preprocessing method for determining the centers of fingerprint images and the neural networks on the performance of an individual identification system. The fingerprint images were classified into four directions, 0°, 45°, 90° and 135°, then two kinds of smoothing were done and the neural networks were optimized. From the results, we found that the preprocessing and the layered neural networks were useful for the individual identification system with a high performance and that the preprocessing method determining the centers of fingerprint images produced higher performance than the system without the preprocessing.

Waveform reconstruction from non-uniform samples with application to speech coding

Abstract: Summary form only given. For non-stationary signals, such as speech, we show that extrema are useful non-uniform samples for a compact reconstruction. We demonstrate the effectiveness of different interpolating functions for reconstructing the signal from the extrema. We also show that such a non-uniform sample based reconstruction scheme can be used for applications like speech coding. In this context, we explore the quantization properties of the extrema information and design a good quality speech coder at 12-13 kbps.