14th European Signal Processing Conference Program

We propose a hierarchical and spatially variant mixture model for image segmentation where the pixel labels are random variables. Distinct smoothness priors are imposed on the доставка цветов в Белгороде probabilities and the model parameters are computed in closed form through maximum a posteriori (MAP) estimation. More specifically, we propose a new prior for the label probabilities that enforces spatial smoothness of different degree for each cluster. By taking into account spatial information, adjacent pixels are more probable to belong to the same cluster (which is intuitively desirable). Also, all of the model parameters are estimated in closed form from the data. The proposed conducted experiments indicate that our approach compares favorably to both standard and previous spatially constrained mixture model-based segmentation techniques.

This article presents a comparative study of 14 supervised evaluation criteria of image segmentation results. A pre-liminary study made on synthetic segmentation results allows us to globally characterise the behaviours of the selected criteria. This first analysis is then completed on a selection of 300 real images extracted from the Corel data-base. Ten segmentation methods based on threshold selection are used to generate real segmentation results and various situations corresponding to under- and over-segmentation. Experimental results permit to reveal the advantages and limitations of the studied criteria face to various situations.

Mean shift is robust for image segmentation through local mode seeking. However, like most segmentation schemes it suffers from over-segmentation due to the lack of semantic information. This paper proposes an enhanced spatial-range mean shift segmentation approach, where over-segmented regions are reduced by exploiting the positions and frequencies at which mean shift filters converge. Based on our observation that edges are related to spatial positions with low mean shift convergence frequencies, merging of over-segmented regions can be guided away from the perceptually important image edges. Simulations have been performed and results have shown that the proposed scheme is able to reduce the over-segmentation while maintaining sharp region boundaries for semantically important objects.

In this paper, we propose a new multi-scale image segmentation relying on a hierarchy of partitions. First of all, we review morphological methods based on connections which produce hierarchical image segmentations and introduce a new way of generating fine segmentations. From connection-based fine partitions, we focus on producing hierarchical segmentations. Multi-scale segmentations obtained by scale-space or region merging approaches have both benefits and drawbacks, therefore we propose to integrate a scale-space approach in the production of a hierarchy of partitions which merges similar regions. Starting from an initial over-segmented fine partition, a region adjacency graph is alternatively simplified and decimated. The simplification of graph nodes models tends to produce similar ones and this is used to merge them. The algorithm can be used to simplify the image at low scales and to segment it at high scales.

Tree-structured Markov random fields have been recently proposed in order to model complex images and to allow for their fast and accurate segmentation. By modeling the image as a tree of regions and subregions, the original K-ary segmentation problem can be recast as a sequence of reduced-dimensionality steps, thus reducing computational complexity and allowing for higher spatial adaptivity. Up to now, only binary tree structures have been considered, which simplifies matters but also introduces an unnecessary constraint. Here we use a more flexible structure, where each node of the tree is allowed to have a different number of children, and also propose a simple technique to estimate such a structure based on the mean shift procedure. Experiments on synthetic images prove the structure estimation procedure to be quite effective, and the ensuing segmentation to be more accurate than in the binary case.

This paper introduces a multi-pronged approach for segmentation of color images of man-made structures like cargo containers, buildings etc. A combination of vector angle computation of the RGB data and the absolute difference between the intensity pixels is used to segment the image. This method has the advantage of removing intensity based edges that occur where the saturation is high, while preserving them in areas of the image where saturation is very low and intensity is high. Thus, unlike previous implementations of vector angle methods and edge detection techniques, relevant objects are better segmented and unnecessary details left out. A novel method for connecting broken edges after segmentation using the Hough transform is also presented.

In the design of the communication protocols for wireless sensor networks a specific requirement emerges from the fact that the data contained in an individual sensor is not important per se, but its significance is instantiated with respect to the contribution to the overall sensing task and the decision fusion. Therefore, the communication protocols should be application-aware and operate by reckoning the utility of the carried data. In this paper we consider the problem of hypothesis testing at the fusion center (sink) when all the sensors communicate with the sink via a random access channel. Each sensor contains a binary information 0 (event occurred) or 1 (event did not occur). In a traditional protocol design, an existing random--access protocol is used by which the sink collects the data from all sensors and subsequently makes the decision through majority voting over the received data. In this paper, we propose approaches for joint design of the communication and the decision fusion for the application of hypothesis testing. The fusion center terminates the data gathering through the random access channel as soon as it can make sufficiently reliable decision based on the data received so far. We describe two instances of the protocols, where the total number of sensors $N$ is known and not known, respectively. Our results show that the proposed approaches provide optimized performance in terms of time, energy and reliability.

We study the role of information feedback for the problem of distributed signal tracking/estimation using a sensor network with a fusion center. Assuming that the fusion center has sufficient energy to reliably feed back its intermediate estimates, we show that the sensors can substantially reduce their power consumption by using the feedback information in a manner similar to the stochastic approximation scheme of Robbins-Monro. For the problem of tracking an autoregressive source or estimating an unknown parameter, we quantify the total achievable power saving (as compared to the distributed schemes with no feedback), and provide numerical simulations to confirm the theoretical analysis.

A wireless sensor network with a fusion center is considered to study the effects of dependent observations on the parameter estimation problem. The sensor observations are corrupted by Gaussian noise with geometric spatial correlation. From an energy point of view, sending all the local data to the fusion center is the most costly, but leads to optimum performance results since all the dependencies are taken into account. From an estimation accuracy point of view, sending only parameter estimates is the least accurate, but is the most parsimonious in terms of communication costs. Hence, this tradeoff between the energy efficiency and the estimation accuracy is explored by comparing the performance of maximum likelihood estimator (MLE) and the sample average estimator (SAE) under various topologies and communication protocols. We start by reviewing the results from the one-dimensional case and continue by extending those results to various two-dimensional topologies. Surprisingly, we discover a class of regular polygon topologies where the MLE under spatial correlation reduces to the SAE.

We consider distributed estimation of a deterministic parameter vector using an ad hoc wireless sensor network. The estimators derived are obtained as solutions of constrained convex optimization problems. Using the method of multipliers in conjunction with a block coordinate descent approach we demonstrate how the resultant algorithms can be decomposed into a set of simpler tasks suitable for distributed implementation. We show that these algorithms have guaranteed convergence to properly defined optimum estimators, and further exemplify their applicability to solving estimation problems where the signal model is completely or partially known at individual sensors. Through numerical experiments we illustrate that our algorithms outperform existing alternatives.

In this work we review some of the most recent in-network computation capabilities that can be used in sensor networks to alleviate the information traffic from the sensors towards the sink nodes. More specifically, after briefly reviewing distributed average consensus techniques, we will concentrate on consensus mechanisms based on self-synchronization of coupled dynamical systems, initialized with local measurements. We will show how to achieve globally optimal distributed detection and estimation through minimum exchange of information between nearby nodes in the case where the whole network observes one common event.

We investigate the spatio-temporal characteristics of the sound field. Spatial sampling using a set of microphones is studied for different array topologies. The reconstruction problem is also discussed. Distributed compression is then addressed using an information-theoretic point of view. In particular, optimal rate-distortion tradeoffs are derived for a linear network setup and a hearing aids configuration.

In powerline or quasi-static wireless systems, the use of a multi-user OFDM based communication is advocated. It is reasonable to consider that the channel is known at the transmitter and receiver. Moreover, in downlink, a spectral mask constraint is usually imposed. Furthermore, to guarantee fairness between all the active users, a balanced rate based criterion optimization is recommended. Therefore, in such previous context, we investigate the achievable rate region in the general case of MC-DS-CDMA, of which the OFDMA is a particular case. Then we propose a simplified algorithm to calculate an approximate balanced rate solution for the OFDMA case. The loss of the OFDMA solution with respect to the MC-DS-CDMA solution is shown to be acceptable. Comparisons with other OFDMA allocation algorithms have also been performed.

This paper presents a low-complexity blind Maximum-Likelihood (ML) detector for Orthogonal Frequency Division Multiplexing (OFDM) systems in block fading channels. We reduce the receiver complexity by subcarrier grouping (SG) in which the OFDM block is partitioned into smaller groups, and then the data are detected on a group-by-group basis. An identifiability analysis is also provided, which shows that the data in each group can be identified under a more relaxed condition than that in [1], enabling us to use small group size for implementation efficiency. Our simulation results show that the proposed detector can provide good symbol error performance even when the group size is much smaller than the discrete Fourier transform size.

Zero-IF based OFDM transmitters and receivers are gaining a lot of interest because of their potential to enable low-cost, low-power and less bulky terminals. However these systems suffer from In-phase/Quadrature-phase (IQ) imbalances in the front-end analog processing which may have a huge impact on the performance. We also consider the case where the local oscillator suffers from carrier frequency offset. As OFDM is very sensitive to the carrier frequency offset, this distortion needs to be taken into account in the derivation and analysis of any IQ imbalance estimation/compensation scheme. In this paper the effect of both transmitter and receiver IQ imbalance under carrier frequency offset in an OFDM system is studied and algorithms are developed to compensate for such distortions in the digital domain.

Based on the estimated channel, this paper presents a new method for first arrival estimation for positioning application in OFDM mobile communication systems. In the new method, the characteristics of the information theoretic criteria is exploited to estimate the time of arrival (TOA). The information theoretic criteria is established on the basic of the different statistical characteristics of noise and the mobile channel. In the proposed algorithm, the calculation of the autocorrelation matrix and their eigenvalues are not required. Therefore, the complexity of the proposed method is low. Simulation results show that the performance of system in terms of the detection rate is very high An accurate estimation of the first arrival path (or the time of arrival) can be obtained even though the first arrival path is strongly attenuated and the system suffers from strong additive noise.

In this paper we propose a frequency-domain IQ-imbalance and carrier frequency offset (CFO) compensation and equalization for OFDM transmission over doubly selective channels. IQ-imbalance and CFO arise due to imperfections in the receiver and/or transmitter analog front-end, whereas user mobility and CFO give rise to channel time-variation. In addition to IQ-imbalance and the channel time-variation, the cyclic prefix (CP) length may be shorter than the channel impulse response length, which in turn gives rise to inter-block interference (IBI). While IQ-imbalance results in a mirroring effect, the channel time-variation results in inter-carrier interference (ICI). The frequency-domain equalizer is proposed to compensate for the IQ-imbalance taking into account ICI and IBI. The frequency-domain equalizer is obtained by transferring a time-domain equalizer to the frequency-domain resulting in the so-called per-tone equalizer (PTEQ).

Orthogonal frequency division multiplexing (OFDM) provides a viable solution to communicate over frequency selective fading channels. However, in the presence of frequency nulls in the channel response, the uncoded OFDM faces serious symbol recovery problems. As an alternative to previously reported error correction techniques in the form of pre-coding for OFDM, we propose the use of post-coding of OFDM symbols in order to achieve frequency diversity. Our proposed novel post-coded OFDM (PC-OFDM) comprises of two steps: 1) upsampling of OFDM symbols and 2) subsequent multiplication of each symbol with unit magnitude complex exponentials. It is important to mention that PC-OFDM introduces redundancy in OFDM symbols while precoded OFDM introduces redundancy in data symbols before performing the IFFT operation. The main advantages of this scheme are reduction in system complexity by having a simple encoder/decoder, smaller size IFFT/FFT (inverse fast Fourier transform/fast Fourier transform) modules, and lower clock rates in the receiver and transmitter leading to lower energy consumption. The proposed system is found to be equally good over Gaussian and fading channels where it achieves the maximum diversity gain of the channel. Simulation results show that PC-OFDM performs better than existing precoded OFDM and Pulse OFDM systems.

A very fast technique for designing discrete wavelet multi-tone (DWMT) transceivers without using time-consuming nonlinear optimization is introduced. In this method, the filters in both the transmitting and receiving filter banks are generated based on the use of a single linear-phase finite-impulse response prototype filter and a cosine-modulation scheme and the prototype filter is optimized by using the windowing technique. The novelty of the proposed technique lies in exploiting spline functions in the transition band of the ideal filter, instead of using the conventional brick-wall filter. In this approach, a simple line search is used for find-ing the passband edge of the ideal filter for minimizing a predetermined cost function. The resulting DWMT trans-ceivers closely satisfy the perfect reconstruction property, as is illustrated by means of examples.

This paper introduces a class of oversampled complex-modulated causal IIR filter banks for flexible frequency-band reallocation networks. In the simplest case, they have near perfect magnitude reconstruction (NPMR), but by adding a phase equalizer they can achieve near-PR.

This paper investigates the mathematical framework of the multiresolution approach under the assumption that the sequence knots are irregularly spaced. The study is based on the construction of nested non-uniform quadratic spline multiresolution spaces. We focus on the construction of suitable quadratic orthonormal spline scaling and wavelet bases. If no more additional conditions than multiresolution ones are imposed, the orthonormal basis of the quadratic spline space is represented, on each bounded interval of the sequence, by three discontinuous scaling functions. Therefore, the quadratic spline wavelet basis, closely related to the scaling basis, is also defined by a set of discontinuous wavelet functions on each bounded interval of the sequence. We show that a judicious orthonormalization procedure of the basic quadratic spline space basis allows to (i) satisfying the continuity conditions of the scaling and wavelet functions, (ii) reducing the number of the wavelet functions to only one function, and (iii) reducing the complexity of the filter bank.

Besides perfect reconstruction and linear phase, regularity is a desirable essential property of filter banks for image coding as it is associated with the smoothness of the related wavelet basis. This paper shows how to constrain quaternionic factorizations of eightband linear phase paraunitary filter banks to have the first regularity structurally imposed. The result is not very general but some facts make it notable. Firstly, these systems are a direct extension of the standard eight-point discrete cosine transform (DCT) and this facilitates practical applications. Secondly, the first regularity eliminates the DC leakage which cause visually annoying checkerboard artifact. Finally, our solution offers clear advantages over the known ones as the regularity conditions are formulated directly in terms of quaternionic lattice coefficients. Namely, both regularity and losslessness can be easily preserved regardless of coefficient quantization unavoidable in finite-precision implementations.

This paper presents a method for designing low-delay nonuniform pseudo QMF banks. The method is motivated by the work of Li, Nguyen and Tantaratana, in which the nonuniform filter bank is realized by combining an appropriate number of adjacent subbands of a uniform pseudo QMF filter bank. In prior work, the prototype filter of the uniform pseudo QMF is constrained to have linear phase and the overall delay associated with the filter bank was often unacceptably large for filter banks with a large number of subbands. By relaxing the linear phase constraints, this paper proposes a pseudo QMF filter bank design technique that significantly reduces the delay. An example that experimentally verifies the capabilities of the design technique is presented.

This paper presents the theory and lattice structures of a large class of oversampled linear phase paraunitary filter banks. We deal with FIR filter banks with real-valued coefficients in which all analysis filters have the same arbitrary filter length and share the same symmetry center. Necessary existence conditions on symmetry polarity of the filter banks are firstly derived. Lattice structures are developed for type-I oversampled linear phase paraunitary filter banks. Furthermore, these lattice structures can be proven to be complete. Finally, several design examples are presented to confirm the validity of the theory and lattice structures.

The linearly constrained minimum variance (LCMV) method is an extension of the classical minimum variance distortionless response (MVDR) filter, allowing for multiple linear constraints. Depending on the spatial filter length and the desired frequency grid, a direct computation of the resulting spatial beampattern may be prohibitive. In this paper, we exploit the rich structure of the LCMV expression to