• Classification de textures fondée sur la theorie des ondelettes hyper-analytiques et les copules, bilateral programme, grant no. 510/31.03.2011, period 2011-2012, partners UPT, IPB-ENSEIRB MATMECA, funded by ANCS,2011-UEFISCDI,2012/EGIDE


    • Modelarea stohastica a imaginilor naturale este obiectul unei game largi de aplicatii: inlaturarea zgomotului, filtrare, clasificare, segmentare, compresie si sinteza imaginii. Scopul este de a avea o descriere statistica cat mai compacta a informatiei din imagine. Principala dificultate care sta in calea modelarii statistice directe a imaginilor naturale este variabilitatea foarte mare a acestora. De aceea se recurge de multe ori la transformarea imaginilor naturale pentru a se obtine reprezentari mai favorabile avand o variabilitate a modelului statistic mai redusa. Proiectarea unui model se confrunta cu problema de dimensionalitate. Natura multivariata a problemei decurge direct din natura datelor achizitionate sau proiectia imaginii in spatii de dimensiuni mai mari (transformare multi-orientare multi-nivel). Prin urmare, manipularea informatiei spatiale necesita lucrul in spatii de mari dimensiuni care pot prezenta printre altele, dependente intre si intra componente. Cercetarile recente in domeniul analizei si sintezei texturilor arata importanta modelarii texturii intr-un domeniu "augmentat" cum este cel al unei transformari multi-orientare multi-nivel (descompunere piramidala). Imaginea se descompune intr-un set de subbenzi orientate; se cauta un model stohastic cu un set minim de parametri (descriptori). Majoritatea metodelor sunt bazate pe un model care implica statisticile marginale. Practic se modeleaza separat fiecare subbanda orientata, prin densitatea sa de probabilitate cu cozi accentuate si se iau in considerare parametrii lor ca descriptori. Daca aceste metode duc la rezultate interesante in diverse aplicatii, cum ar fi de clasificare sau de filtrare, in sinteza raman foarte limitate ca performanta. intr-adevar, acest model are doua dezavantaje majore: 1) se bazeaza pe ipoteza de independenta intre componente, 2) nu foloseste statisticile mutuale pentru a tine cont de fiecare componenta. in cadrul acestui proiect, cautam modele adecvate compacte, impreuna cu analiza si sinteza de imagini texturate. Obiectivul general al proiectului este imbunatatirea metodelor existente de clasificare a texturilor, folosind abordari statistice in domeniul unor transformate complexe, si anume transformata wavelet hiperanalitica. Unul dintre punctele forte ale studiului nostru este propunerea unei abordari ce foloseste o noua transformata cu selectivitate directionala crescuta si cvasi-invarianta la translatii, si anume transformata wavelet hiperanalitica (HWT) propusa de echipa romana precum si o noua familie de modele stohastice bazate pe modelul stohastic multivariat al copulelor dezvoltate de echipa franceza. Pe baza celor doua propuneri, obiectivul nostru este dezvoltarea unei metode de clasificare a texturilor, eficienta si robusta, organizata in trei etape si anume 1) aplicarea unei transformari, 2) estimarea parametrilor complecsi ai modelului stohastic, 3) evaluarea unei masuri de similaritate intre texturi folosind acesti parametri complecsi. in plan experimental, vom aborda in special cazul imaginilor de textura "prelevate" in conditii variabile (luminanta, unghi de rotire, dimensiune etc). in acest proiect ne propunem sa optimizam pasii 2 si 3 abordand in special modelarea interdependentei coeficientilor transformarii intra/inter nivel. Metodele propuse sunt analizate, si vor fi selectionate cele mai bune solutii pornind de la baze de date de texturi precum VISTEX si OUTEX.
    Stochastic modeling of natural images is subject to a wide range of applications: denoising, filtering, classification, segmentation, image compression and synthesis. The goal is to have a compact statistical description of the information contained in the image. The main difficulty in the way of direct statistical modeling of natural images is their high variability. Therefore for natural images, a transform is often used to obtain a more favorable representation of the statistical model with a lower variability. Designing a good model faces the problem of dimensionality. The multivariate nature of the problem comes directly from the nature of the data or its projection on higher dimension spaces (multi-level multi-orientation transform). Therefore, manipulation of spatial information requires working with high dimension spaces that may have, among others, and intra- and inter-dependencies between components. Recent research in the analysis and synthesis of texture shows the importance of modeling texture in an "augmented" field like that of a multi-level multi-orientation transform (pyramid decomposition). The image is decomposed into a set of oriented subbands; a stochastic model with a minimal set of parameters (descriptors) is searched. Most methods are based on a model based on marginal statistics. Each oriented subband is modeled separately by its probability density with heavy tails and its parameters are considered as descriptors. These methods lead to interesting results in various applications, such as classification or filtering, but in synthesis they are very limited in performance. Indeed, this model has two major drawbacks: 1) it is based on the assumption of independence between components, 2) it does not use mutual statistics to account for each component. In this project, we seek suitable compact models, together with the analysis and synthesis of texture images. The overall objective is to improve existing methods of classification of textures using statistical approaches in the field of complex transforms, namely the hyperanalytic wavelet transform. One of the strengths of our study is to propose an approach that uses a new transform with increased directional selectivity and quasi-invariance to translations, namely the hyperanalytic wavelet transform (HWT) proposed by the Romanian team and a new family of stochastic models based on the multivariate copula model developed by the French team. Based on the two proposals, our goal is to develop a method for classifying textures, efficient and robust, organized in three stages namely 1) applying a transform, 2) estimation of complex parameters for the stochastic model, 3) evaluation of a measure of similarity between textures using these complex parameters. In an experimental plan, we will address in particular the case of textured images "taken" under varying conditions (light, angle of rotation, size, etc.). In this project we intend to optimize steps 2 and 3 addressing in particular the modeling of intra / inter level interdependence of coefficients. The proposed methods are analyzed, and the best solutions will be selected from texture databases such as VISTEX and OUTEX.
    Visits: 2011
    14.07-23.07.2011 mobility France-Romania, Turcu Flavius, Turcu Serban Ioana
    5.11-11.11.2011 mobility Romania-France, Nafornita Corina, Isar Alexandru, Nafornita Ioan
    2012
    August 2012, mobility France-Romania, Yannick Berthoumieu, conferinta EUSIPCO2012
    15.10-21.10.2012 mobility Romania-France, Nafornita Corina, Isar Alexandru, Nafornita Ioan
    Papers: ISI Journals
    1.Isar Alexandru, Space-Frequency Localization As Bivariate Mother Wavelets Selecting Criterion for Hyperanalytic Bayesian Image Denoising, Fluctuations and Noise Letters, vol. 11, 2012, 8 pages.
    2. N. E. Lasmar, Y. Berthoumieu, Gaussian Copula Multivariate Modeling for Image Texture Retrieval Using Wavelet Transforms, IEEE Transactions on Image Processing 99, 2012
    3. Abdourrahmane M. Atto, Yannick Berthoumieu: Wavelet Packets of Nonstationary Random Processes: Contributing Factors for Stationarity and Decorrelation. IEEE Transactions on Information Theory 58(1): 317-330 (2012)
    4. Firoiu, I. ; Nafornita, C. ; Isar, D. ; Isar, A. ; Bayesian Hyperanalytic Denoising of SONAR Images, IEEE Geoscience and Remote Sensing Letters, vol. 8, Issue 6, 2011, pp. 1065-1069,

    Book chapters
    1. Corina Nafornita, Alexandru Isar, Application of Discrete Wavelet Transform in Watermarking-chapter 12 in "Discrete Wavelet Transforms. Algorithms and Applications", Edited by Hannu Olkkone, INTECH, 2011, ISBN 978-953-307-482-5, 197-218.
    2. Alexandru Isar, Ioana Firoiu, Corina Nafornita, Sorin Moga, Sonar Images Denoising-chapter 8 in "Sonar Systems", Edited by Nikolai Kolev, INTECH Janeza Trdine 9, 51000, Rijeka, Croatia, 2011, ISBN 978-953-307-345-3, 173-206.

    Conference Proceedings
    1. Corina Nafornita, Alexandru Isar, "A Complete Second Order Statistical Analysis of the HWT", Int. Symp on Electronics and Telecommunications, ISETC2012, Timisoara 15-16 nov 2012. in press 2. Corina Nafornita, Alexandru Isar, Ioan Nafornita, The Hyperanalytic Wavelet Packets -A Solution to Increase the Directional Selectivity in Image Analysis, Int. Symp on Electronics and Telecommunications, ISETC2012, Timisoara 15-16 nov 2012. in press
    3. Aurelien Schutz, Yannick Berthoumieu, Flavius Turcu, Corina Nafornita, Alexandru Isar, Barycentric Distribution Estimation For Texture Clustering Based On Information-Geometry Tools, Int. Symp on Electronics and Telecommunications, ISETC2012, Timisoara 15-16 nov 2012. in press
    4. Corina Nafornita, Yannick Berthoumieu, Ioan Nafornita, Isar Alexandru, Kullback-Leibler Distance Between Complex Generalized Gaussian Distributions, Proceedings of International Conference Eusipco 2012, Bucuresti, Romania, August 27-31, 2012, ISBN: 778-2-3234-7919-1, IEEE Xplore,
    5. Ahmed Drissi El Maliani, Mohammed El Hassouni, Yannick Berthoumieu, Driss Aboutajdine: Multi-model Approach for Multicomponent Texture Classification. ICISP 2012: 36-44
    6 Atto, A.M.; Berthoumieu, Y.; How to perform texture recognition from stochastic modeling in the wavelet domain, Acoustics, Speech and Signal Processing (ICASSP), 2011 IEEE International Conference on, 4320-4323, Prague,
    7. Atto, A.M.; Berthoumieu, Y.; Structuring of large and heterogeneous texture databases, Statistical Signal Processing Workshop (SSP), 2011 IEEE, 28-30 June 2011, 589-592, Nice,
    8. Corina Nafornita, Dorina Isar, Alexandru Isar, Searching the Most Appropriate Mother Wavelets for Bayesian Denoising of Sonar Images in the Hyperanalytic Wavelet Domain, Proceedings of the IEEE Workshop on Statistical Signal Processing SSP 2011, 28-30 June, 2011, Nice, France, ISBN 978-1-4577-0568-7, pp. 169-172,
    9. Beatrice Arvinti-Costache, Marius Costache, Corina Nafornita, Alexandru Isar, Ronny Stoltz, Hannes Toepfer, A Wavelet Based Baseline Drift Correction Method for Fetal Magnetocardiograms, Proceedings of the 9th IEEE International NEWCAS Conference, June 26-29, 2011, Bordeaux, France,

Intelligent Signal Processing Centre
Copyright © 2010