Felix Totir

Hierarchical Segmentation of Polarimetric SAR Images Using Heterogeneous Clutter Models

In this paper, heterogeneous clutter models are used to describe polarimetric synthetic aperture radar (PolSAR) data. The KummerU distribution is introduced to model the PolSAR clutter. Then, a detailed analysis is carried out to evaluate the potential of this new multivariate distribution. It is implemented in a hierarchical maximum likelihood segmentation algorithm. The segmentation results are shown on both synthetic and high-resolution PolSAR data at the X- and L-bands. Finally, some methods are examined to determine automatically the “optimal” number of segments in the final partition.

Superresolution algorithms for spatial extended scattering centers

Scattering centers model (SCM) is usually considered for modeling target backscattered signal in high-resolution radar. In this case the impulse response of each scattering center is represented by a time-delayed Dirac pulse. Some of most popular superresolution imagery techniques, such as MUSIC or ESPRIT, are well-matched to this model. Under this hypothesis, they outperform Fourier-based techniques in terms of both spatial and dynamic resolutions. However, the behavior of real-world targets is often very different from that of the SCM. Indeed, their reflectivity function is not confined just to several perfectly localized scattering centers, but it can be rather approximated by a set of scattering regions having different spatial extent. SCM becomes then inappropriate and the superresolution methods may provide unexpected results. Furthermore, the amplitude information is difficult to interpret in this case. In this paper we propose an extension of two superresolution methods, MUSIC and ESPRIT, to cope with extended scattering centers (ESC). According to this model, the impulse response of an ESC is not a Dirac pulse, but a window of finite support. Besides the position, the size (spatial extent) of this window is also recovered. This additional information about the target structure can be used for increasing ATR (automatic target recognition) accuracy and robustness.

Supervised self-organizing classification of superresolution ISAR images: an anechoic chamber experiment

The problem of the automatic classification of superresolution ISAR images is addressed in the paper. We describe an anechoic chamber experiment involving ten-scale-reduced aircraft models. The radar images of these targets are reconstructed using MUSIC-2D (multiple signal classification) method coupled with two additional processing steps: phase unwrapping and symmetry enhancement. A feature vector is then proposed including Fourier descriptors and moment invariants, which are calculated from the target shape and the scattering center distribution extracted from each reconstructed image. The classification is finally performed by a new self-organizing neural network called SART (supervised ART), which is compared to two standard classifiers, MLP (multilayer perceptron) and fuzzy KNN ( nearest neighbors). While the classification accuracy is similar, SART is shown to outperform the two other classifiers in terms of training speed and classification speed, especially for large databases. It is also easier to use since it does not require any input parameter related to its structure.

PolSAR images characterization through Blind Sources Separation techniques

Since the backscattered signal in PolSAR images is intrinsically linked with the physical characteristics of the objects in the image, valuable information may be extracted therefrom. The paper focus is to propose a new physical characterization of the scattering target, inspired by the Blind Sources Separation techniques.

A TD-PSOLA based method for speech synthesis and compression

Mobility and cost restrictions of current text-to-speech systems stop them from being used by people with speech impairments all over the world. Therefore new ways to improve mobility and lower cost have to be developed. This can be done by decreasing the computational resources used by speech synthesis systems. Non-parametric concatenative synthesis techniques provide the easiest way to generate artificial speech with high quality. Although, they can be, in general, computationally efficient (e.g., TD-PSOLA) they are not always suited for implementation on embedded devices because they require rather large recorded speech data-bases. A big part of the recorded speech data is represented by the samples of the vowels. Therefore, compression ratios of at least 25% can be achieved for Romanian, by removing all these samples but one overlap-add (OLA) frame. At synthesis, the remaining vowel is used to generate the original sound. The paper presents a new method for the generation and the compression of vowels, starting from only one OLA frame and using TD-PSOLA in new way. Experiments show that by appropriately choosing pitch and amplitude jitter models, high quality synthetic speech can be achieved.

On letter to sound conversion for Romanian: A comparison of five algorithms

This paper presents an evaluation of 5 letter-to-sound (LTS) systems for Romanian. The first is an expert system; three of them use automatic classification methods with decision trees, neural networks and support vector machines respectively and the fifth one uses pronunciation by analogy. All systems were trained and tested on the same database: a 15,517 words corpus built according to the SpeechDat specifications and a corpus consisting of the most frequent 4779 words in Romanian. The results show that decision trees and neural networks generate the best results for letter to sound conversion in Romanian.

ICA-based information extraction method for PolSAR images

Identifying and characterizing objects in PolSAR images rely on proper backscattered signal models, correct segmentation algorithms and on exploiting the connections between the phisical structure of the considered objects and their radar echos. The paper presents an ICA-based approach for characterizing physical properties of the objects on the Earth surface. Segmentation of PolSAR image is based on stochastic models. Each segment is assimilated to an object or surface type and pixels of the segment are treated like realizations of the underlying target scattering mechanism. This scattering mechanism is characterized through ICA-based approach, thus providing information about the corresponding target. The TSVM decomposition is used to analyze the mixing matrix issued from ICA and to link its values to physical properties of the target.

Circularity of complex stochastic models in PolSAR and multi-pass InSAR images

Polarimetry and multi-pass interferometry extend the dimensionality of SAR data, so the necessity to have multivariate statistic (and non-Gaussian, because of the high resolution) distributions as models for these types of data: such are the SIRV (Spherically Invariant Random Vectors). However, as the statistic model becomes so complicated, correctly estimating its parameters gets difficult. More, although they are versatile, the SIRV models are not guaranteed to match the PolSAR / InSAR data. To evaluate the pertinence of those models with respect to the PolSAR data, through one of their most important statistic property, namely the circularity, it is the purpose of this paper.

Influence of the magnitude and phase errors on the reconstructed radar images

Our research work has been focused for several years on ISAR techniques and ATR using superresolution radar imagery. The anechoic chamber of ENSIETA at Brest and the associated measurement facilities allow us to obtain radar signatures for various scale reduced targets and to reconstruct their radar images using a turntable configuration. The paper investigates an aspect which has a great influence on the quality of the reconstructed images. We use two canonical targets, a single sphere and a set five spheres, in order to have perfect control of the target configuration and to make the interpretation of the obtained results easier. The idea of the paper is quite simple and was suggested by some errors in the scattering center positioning we detected on the reconstructed images. Hence, we decided to investigate the influence of the magnitude and the phase of acquired signatures on the scattering center position estimation.

Multidimensional superresolution ISAR reconstruction techniques in sea-cluttered environnement

For high-resolution radars (as the ones used for radar imagery tasks are), sea clutter is difficult to characterize, especially for low grazing angles. According to the physical model of sea surface, the echo signal is best described as the product of a Gaussian process, with rapid variation (the speckle component), and of a slow varying component (the underlying component). The K-compound distribution is used for the statistical characterization of the clutter echo. Standard target complex signature simulations techniques (based on the well-known scattering center model or using geometric primitives) are combined with synthetic sea clutter, allowing us to generate realistic signals. The MUSIC 2D super-resolution reconstruction method, presented in the paper, is used to generate ISAR target images. A comparative study of robustness of MUSIC and Fourier type images is then performed. The obtained results show the validity of both clutter simulation engine and super-resolution reconstruction methods. The problem of naval targets, subject to particular simulation conditions, is also discussed in detail.