Texture classification

Top PDF Texture classification:

Fuzzy aura matrices for texture classification

Fuzzy aura matrices for texture classification

very small descriptors (p = 4). FGLAMs also compare favorably with rotation-invariant CLBPs on the dataset TC-20, which shows that FGLAMs inherently handle rotation variations quite well. 5.3.2. Comparison with other fuzzy texture descriptors We also compare the performances reached by FGLAMs with those provided by other fuzzy texture descriptors. The Fuzzy LBPs (FLBPs) proposed by Keramidas et al. [ 19 ] make each pixel’s neighbor- hood partly contribute to several bins of the LBP histogram. The Local Fuzzy Patterns (LFPs) of Vieira et al. [ 38 ] represent each pixel’s neighborhood as a fuzzy set defined by the membership degree of neighboring pixels to the central value. The sigmoid is used as membership function, which generalizes the step function of “crisp” LBPs. We test these two approaches on the Outex gray-level and color datasets with the optimal parameters given by their authors, but with the 1NN classification scheme as before. In order to obtain an equal size of 256 for FLBP and LFP histograms and for FGLAMs so that these three descriptors are fairly compared, we compute FGLAMs by setting p = 16 and FLBPs with a 8-neighborhood instead of the full 3 × 3 neighborhood of [ 19 ]. Table 5(c) shows that LFPs are somewhat robust to rotation but are in general of minor interest. FLBP is a far more robust descriptor that notably shows superiority with regard to FGLAMs when the image resolution changes. However, FGLAMs perform better on color images. At last, we compare FGLAM performance with that of the Fuzzy Co-occurrence Matrices (FzCMs) developed by Munklang et al. [ 23 ]. This approach uses the fuzzy C-means algorithm to quantize the gray levels, then builds eight fuzzy GLCMs for each main direction of the image plane, and extracts the average and standard deviations of four Haralick’s features from these matrices. For a fair comparison with results in [ 23 ], we implement the classifi- cation scheme based on the one-versus-all strategy of a multi-class Support Vector Machine (SVM) with σ = 0.25 in the radial basis function kernel and ǫ = 10 −3 as termination criterion tolerance, and we similarly validate the classification thanks to a 10-fold cross validation. We compare the texture classification results on the (challenging) UIUC dataset that contains 25 texture classes of 40 gray-level images [ 39 ]. Because the best classification accuracy of FzCMs (77.0%) is reached with 64-dimensional feature vectors [ 23 ], we compute FGLAMs of size 64 by setting p = 8. As FGLAM c and FGLM l provide accuracies of 82.7% and 79.9%, FGLAMs outperform FzCMs with a SVM classifier.
En savoir plus

38 En savoir plus

Impact of topology-related attributes from Local Binary Patterns on texture classification

Impact of topology-related attributes from Local Binary Patterns on texture classification

tors like LBP-rotation invariant uniform patterns (LBP riu2 ), and Local Binary Count (LBC). Like them, it allows contrast and rotation invari- ant image description using more compact descriptors than classic LBP. However, its expressiveness, and then its discrimination capability, is higher, since it includes additional information, including the number of connected components. The impact of the different attributes on texture classification performance is assessed through a systematic comparative evaluation, performed on three texture datasets. The results validate the interest of the proposed approach, by showing that some combinations of attributes outperform state-of-the-art LBP-based texture descriptors.
En savoir plus

15 En savoir plus

RGBD object recognition and visual texture classification for indoor semantic mapping

RGBD object recognition and visual texture classification for indoor semantic mapping

Ground and wall type classification is required in the competition and is also necessary to avoid gravel areas that our robot cannot cross. We chose to use vision for this task, reducing the problem to that of visual texture classification. Many methods for texture classification have been proposed relying on filter banks. These filters encode the local spatial variation that characterize a texture (e.g., [18], [19]) and are used as pre-processing for a classification algorithm such as Support Vector Machine of Nearest Neighbour. However computing such filter may be computationally expensive and other approaches have been shown to give better performances by directly processing image patches [20]. In particular, the approach proposed by [21] relies on randomized trees applied directly to random image sub-windows without any pre- processing and provides very good performances at a small computational cost.
En savoir plus

7 En savoir plus

Texture classification of photographic papers: improving spectral clustering using filterbanks on graphs

Texture classification of photographic papers: improving spectral clustering using filterbanks on graphs

3 Paul Messier LLC, Boston, USA {prenom.nom}@ens-lyon.fr; herwig.wendt@irit.fr; admin@paulmessier.com R´esum´e – Du point de vue du traitement du signal sur graphe, la m´ethode classique de classification dite de “clustering spectral” apparaˆıt comme un banc de filtres passe-bas id´eal. Int´egrant une adaptation de la d´etection multi´echelle de communaut´es de [11] au concept de cœurs de communaut´es [8], nous proposons une m´ethode de classification bas´ee sur d’autres bancs de filtres plus adapt´es aux donn´ees. Dans le cadre d’une classification de textures de papiers photos utile en histoire de l’art, les r´esultats de cette m´ethode s’av`erent plus riches et ais´ement interpr´etables.
En savoir plus

5 En savoir plus

Spatio-spectral binary patterns based on multispectral filter arrays for texture classification

Spatio-spectral binary patterns based on multispectral filter arrays for texture classification

OCIS codes: (110.4234) Multispectral and hyperspectral imaging; (100.0100) Image processing. http://dx.doi.org/10.1364/ao.XX.XXXXXX 1. INTRODUCTION Among multispectral imaging technologies, “linescan” devices acquire frames of narrow spatial slices for several spectral bands then, once the whole scene has been scanned, form the mul- tispectral image by stacking the acquired frames. “Snapshot” devices can oppositely acquire the multispectral image from a single shot. Multi-sensor snapshot systems straight form a fully-defined multispectral image thanks to dichroic beam split- ters that selectively redirect the incoming light by wavelength onto the sensors. Because these systems are expensive and are sensitive to a limited number of bands, single-sensor snapshot systems have been recently developed [ 1 , 2 ]. Most of them use a multispectral filter array (MSFA) laid over the sensor that spec- trally samples the incident light, like the widely-used Bayer color filter array (CFA) in color imaging. The MSFA is defined by a basic periodic pattern in which each filter is sensitive to a narrow spectral band. Each pixel of the resulting raw image is then characterized by one single band according to the MSFA. Such technology achieves a compromise between spatial and spectral samplings (see Sec. A ). The fully-defined multispectral image is estimated by a demosaicing process (see Sec. B ). This process may be greedy and is prone to generate spatio-spectral estimation artifacts. We therefore propose to avoid it for tex- ture classification and design a texture descriptor that is directly computed from the raw image (see Sec. C ).
En savoir plus

11 En savoir plus

Discrete wavelet for multifractal texture classification: application to medical ultrasound imaging

Discrete wavelet for multifractal texture classification: application to medical ultrasound imaging

The proposed method establishes a multifractal analysis framework of such images based on a new multiresolution indicator, called the maximum wavelet coefficient, derived [r]

5 En savoir plus

Unsupervised amplitude and texture classification of SAR images with multinomial latent model

Unsupervised amplitude and texture classification of SAR images with multinomial latent model

segmentation [31] as a robust statistical model. The secondary target in land cover classification from SAR images is to find spatially connected and smooth class label maps. To obtain smooth and segmented class label maps, a post-processing can be applied to roughly classified pixels, but a Bayesian approach allows us to include smoothing constraints to the classification problems. Potts-Markov image model is introduced in [33] for discrete intensity images. In [34] and [35], some Bayesian approaches are exploited for SAR image segmentation. Hidden Markov chains and random fields are used in [36] for radar image classification. [37] exploits a Potts-Markov model with MnL class densities in hyperspectral image segmentation. A double MRFs model is proposed in [23] for optical images to model the texture and the class labels as two different random fields. In [38], am- plitude and texture characteristics are used in two successive and independent schemes for SAR multipolarization image segmentation. In our spatial smoothness model, we assign a binary class map for each class which indicates the pixels belonging to that class. We introduce the spatial interaction within each binary map adopting multinomial logistic model [39]. In our logistic regression model, the probability of a pixel label is proportional to the linear combination of the surrounding binary pixels. If we compare the Potts-Markov image model [33] with ours, we may say that we have K different probability density functions for each binary random field, instead of a single Gibbs distribution defined over a multi-level label map. The final density of the class labels is constituted by combining K probability densities into a multinomial density.
En savoir plus

13 En savoir plus

Texture Classification with Generalized Fourier Descriptors in Dimensionality Reduction Context: an Overview Exploration

Texture Classification with Generalized Fourier Descriptors in Dimensionality Reduction Context: an Overview Exploration

the average error rate for each classification method, we performed 20 iterative experiments with a 10-fold cross validation procedure. 3 Dimensionality Reduction Methods The GFD provide features that are of great potential in pattern recognition as it was shown by Smach et al. in [8]. Unfortunately, these high dimensional datasets are however difficult to handle, the information is often redundant and highly correlated with one another. Moreover, data are also typically large, and the computational cost of elaborate data processing tasks may be prohibitive. Thus, to improve the classification performance it is well interesting to use Dimensionality Reduction (DR) techniques in order to transform high-dimensional data into a meaningful representation of reduced dimensionality. At this time of our work, we selected a dozen of DR methods. However, it is important to note that works employing recent approaches as it could be find in [18] are being finalized (another distance, topology or angle preservation methods like Kernel Discriminant Analysis, Generative Topographic Mapping, Isotop, Conformal Eigenmaps,…).
En savoir plus

12 En savoir plus

Feature Extraction on Local Jet Space for Texture Classification

Feature Extraction on Local Jet Space for Texture Classification

The performance of the proposal is assessed over four databases: Bro- datz [12], Usptex [13], Vistex [14], and Outex [15]; using four texture feature descriptors: Fourier, Gabor, Local Binary Pattern and Local Binary Pat- tern Variance. A principal component analysis (PCA) [16] was performed to reduce the dimension of features and linear discriminant analysis (LDA) is used to perform the classification. In most experiments, the proposed approach obtained higher success rate compared to the feature descriptors applied without the local jet decomposition.

21 En savoir plus

Analyse de la texture des images mammaires par une fusion des lois de Zipf et des Ondelettes pour la classification des tumeurs mammaires via l’analyse en composantes principale

Analyse de la texture des images mammaires par une fusion des lois de Zipf et des Ondelettes pour la classification des tumeurs mammaires via l’analyse en composantes principale

Nous distinguons 4 classes (modèles) de textures (Tuceryan et al.1998) : • Les méthodes statistiques Les méthodes statistiques étudient les relations entre un pixel et ses voisins et définissent des paramètres discriminants de la texture en se basant sur des outils statistiques. Ces modèles statistiques sont efficaces pour de nombreuses textures naturelles ayant des primitives discernables ainsi que pour la caractérisation des structures fines, sans régularité apparente. Notons que plus l'ordre de la statistique est élevé et plus le nombre de pixels mis en jeu est important. Parmi ces méthodes, citons la méthode basée sur les matrices de cooccurrences, celle des matrices de longueurs de plage ( Bagadi A.2016) . La méthode basée sur les matrices de cooccurrence, proposée par Haralick (Haralick. 1979), constitue à explorer les dépendances spatiales des pixels en construisant d’abord une matrice de cooccurrence basée sur l’orientation et la distance entre les pixels de l’image. A partir de ces matrices, nous pouvons extraire des caractéristiques de la texture, comme le contraste, l’entropie ou la différence inverse des moments. Ces modèles sont efficaces pour de nombreuses textures naturelles qui ont des primitives discernables ainsi que pour la caractérisation des structures fines. Plus l’ordre de la statistique est élevé et plus le nombre de pixels mis en jeu est important (Hamoud M.2015).
En savoir plus

71 En savoir plus

Multiscale Anisotropic Texture Analysis and Classification of Photographic Prints: Art scholarship meets image processing algorithms

Multiscale Anisotropic Texture Analysis and Classification of Photographic Prints: Art scholarship meets image processing algorithms

and 10% quantiles). Interestingly, the examination of print thickness, highlight color, reflectance, and texture had led to the conclusion that prints of these pairs were similar. For prints approaching a distance close to the distribution median dis- tance (the solid vertical line in Figure 10), the similarities are less clear. In all respects (texture, reflectance, highlight color, and print thickness), pairs in this range by Max Burchartz (Ger- man, 1887–1961), John Gutmann (American, 1905–1998), and Jaromir Funke (Czech, 1896–1945) are classified as being on the same paper. However, pairs in this same range by Edmund Collein (German, 1906–1992), Helmar Lerski (Swiss, 1871– 1956), and another Gutmann print are categorized as being on different paper when gloss, highlight color, and paper thickness are taken into account. In particular, the prints by Paul Citroen (Dutch, 1896–1983), though they have low texture differences (31% quantile), are classified as being on different paper mainly due to significant difference in gloss and print thick- ness. These results are not surprising given the common manufacturer practice of applying the same texture to different papers. These results suggest that an automated solution for discriminating material-based affinities across collections can- not rely on a single criterion, such as texture or reflectance, for determining results especially for distances around the median of the distribution.
En savoir plus

11 En savoir plus

TEXTURE REMOVAL BY PIXEL CLASSIFICATION USING A ROTATING FILTER

TEXTURE REMOVAL BY PIXEL CLASSIFICATION USING A ROTATING FILTER

5 degrees. Fig. 5(c) shows flat areas results : a black pixel matches at least one flat area, whereas white pixels denote that no flat area has been detected. In others words, they correspond to a pixel lying between two regions (contours), between a homogeneous region and texture region or a cor- ner. The result of the anisotropic diffusion is presented in the Fig. 5(d) after 50 iterations. Note that different objects are perfectly visible whereas textures regions are smoothed and some of them have merged.

5 En savoir plus

Analyse de la texture des images mammaires par une fusion des lois de Zipf et des SFTA pour la classification des tumeurs mammaires via l’analyse en composantes principales

Analyse de la texture des images mammaires par une fusion des lois de Zipf et des SFTA pour la classification des tumeurs mammaires via l’analyse en composantes principales

33 Pour analyser la texture du tissu mammaire avec les SFTA (Segmentation-based Fractal Texture Analysis), les descripteurs sont extraits à partir des zones d'intérêt (ROI : Region Of Interest) segmentées sur les mammographies en les décomposant en un ensemble d'images binaires à partir duquel les dimensions fractales sont calculées [51]. Ces dimensions sont utilisées pour décrire les motifs de texture segmentés des mammographies. Pour la décomposition des zones d'intérêt (ROIs) d'échelle de gris en entrée, la méthode de Two- Threshold Binary Decomposition (TTBD) est appliquée. Dans ce type de décomposition, un ensemble de nt nombre de valeurs de seuil de région d’intérêt en échelle de gris est calculé en appliquant récursivement l'algorithme Ostu multi niveau, où nt est le paramètre défini par l'utilisateur (nous avons procédé par expérimentation et nt=4 nous a semblé la valeur optimale, en effet, lorsque la valeur de nt est plus grande, le nombre des descripteurs redondants augmente et par conséquent la complexité et les besoins en mémoire augmentent aussi). Un ensemble de zones d'intérêt binaires (bROI) est calculé à partir d'une ROI à échelle de gris en appliquant une Two-Threshold Binary Decomposition (TTBD). Dans cette méthode, des paires
En savoir plus

75 En savoir plus

analyse de la texture des images mammaires par une fusion des lois de Zipf et desLBP pour la classification des tumeurs mammaires via les algorithmes génétiques

analyse de la texture des images mammaires par une fusion des lois de Zipf et desLBP pour la classification des tumeurs mammaires via les algorithmes génétiques

Chapitre IV Aide au diagnostic du Cancer du sein assisté par ordinateur médical du cancer du sein (CAD) soit basée sur une large gamme de caractéristiques de texture ainsi que de fonctionnalités obtenues avec d'autres méthodes de traitement d'image. Dans [33] les auteurs affirment que l'analyse manuelle d'image mammaire par un spécialiste est très compliquée du fait du faible contraste. Comme alternative, les auteurs proposent le diagnostic assisté par ordinateur du tissu mammaire basé sur l’analyse de la texture de toute la région de l’image, en partant des régions externes de l’image pour s’étendre aux régions intérieures. Pour cette tâche, ils utilisent des indices de diversité phylogénétique qui désigne une branche de la génétique traitant des modifications génétiques au sein des espèces animales ou végétales : la diversité pure ou la diversité phylogénétique; la somme des distances phylogénétiques; la distance moyenne du voisin le plus proche; la variabilité des espèces phylogénétiques; et la richesse en espèces phylogénétiques.
En savoir plus

75 En savoir plus

Unsupervised amplitude and texture based classification of SAR images with multinomial latent model

Unsupervised amplitude and texture based classification of SAR images with multinomial latent model

Equipe-Projet Ariana Rapport de recherche n° 7700 — version 2 — initial version July 2011 — revised version May 2012 — 29 pages Abstract: We combine both amplitude and texture statistics of the Synthetic Aperture Radar (SAR) images for classification purpose. We use Nakagami density to model the class amplitudes and a non-Gaussian Markov Random Field (MRF) texture model with t-distributed regression error to model the textures of the classes. A non-stationary Multinomial Logistic (MnL) latent class label model is used as a mixture density to obtain spatially smooth class segments. The Classification Expectation-Maximization (CEM) algorithm is performed to estimate the class parameters and to classify the pixels. We resort to Integrated Classification Likelihood (ICL) criterion to determine the number of classes in the model. We obtained some classification results of water, land and urban areas in both supervised and unsupervised cases on TerraSAR-X, as well as COSMO-SkyMed data.
En savoir plus

33 En savoir plus

Neural Approach for Context Scene Image Classification based on Geometric, Texture and Color Information

Neural Approach for Context Scene Image Classification based on Geometric, Texture and Color Information

{ameni.sessi.tn,wael.ouarda,chokri.benamar}@ieee.org, serge.miguet@univ-lyon2.fr Abstract. Revealing the context of a scene from low-level features rep- resentation, is a challenging task for quite a long time. The classification of landscapes scenes to urban and rural categories is a preliminary task for landscapes scenes understanding. Having a global idea about the scene context (rural or urban) before investigating its details, would be an interesting way to predict the content of that scene. In this paper, we propose a novel features representation based on skyline, colour and texture, transformed by a sparse coding using Stacked Auto-Encoder. To evaluate our proposed approach; we construct a new database called SKYLINEScene Database containing 2000 images of rural and urban landscapes with a high degree of diversity. Many experiments were car- ried out using this database. Our approach shows it robustness in land- scapes scenes classification.
En savoir plus

11 En savoir plus

MRI Texture-Based Classification of Dystrophic Muscles. A Search for the Most Discriminative Tissue Descriptors

MRI Texture-Based Classification of Dystrophic Muscles. A Search for the Most Discriminative Tissue Descriptors

The same database of images was analyzed in our previous work [9], as both studies are part of the European COST Action BM1304 project (“MYO-MRI”) aimed at exploring strategies for muscle imaging texture analysis. We differenti- ated between GRMD and healthy dogs at each of three phases of canine growth. Textural features were extracted using statistical, model-based, and filter-based TA methods. Eight sets of features, each derived from a different method, were considered separately. The set of all features derived from all methods was also tested. Experiments involving five classifiers showed that highly satisfactory clas- sification results can be obtained with certain (relatively small) sets of features, especially those based on RLM and co-occurrence matrices (COM) [15]. The work did not perform any feature selection nor attempt to differentiate between tissues in different phases of dystrophy development.
En savoir plus

13 En savoir plus

SAR Image Classification with Non-stationary Multinomial Logistic Mixture of Amplitude and Texture Densities

SAR Image Classification with Non-stationary Multinomial Logistic Mixture of Amplitude and Texture Densities

There are some methods to combine the results of dif- ferent classifiers [8]. Rather than this, we construct a single classifier using the Products of Experts (PoE) approach [9] to combine both SAR amplitude and texture features into a fi- nite mixture model. In our mixture model, we assume that each latent class label is a categorical random variable which is a special version of the multinomial random variable where each pixel belongs to only one class. For each class, we have a binary map that indicates the pixels belonging to that class. We introduce the spatial interaction of each binary map using multinomial logistic model [10] to obtain a smooth segmenta- tion map. Note that the edge preserving segmentation is out of the scope of this paper. In this logistic regression model, the probability of the class label is proportional to a linear com- bination of surrounding binary pixels. In contrast to Potts- Markov image model [11], we have K different probability density functions for random fields of each class, instead of a single Gibbs distribution. The final density of the class la- bel is constituted by combining K probability densities into a multinomial density. In this way, we obtain a non-stationary multinomial class density function which incorporates both class mixture probabilities and spatial smoothness into a sin- gle density. A non-stationary finite mixtures model been has introduced for image classification in [12]. A single model and algorithm are preferred to avoid the propagation of the error between different models and algorithms.
En savoir plus

5 En savoir plus

Texture-adaptive mother wavelet selection for texture analysis

Texture-adaptive mother wavelet selection for texture analysis

Wavelets-based approaches for texture analysis usually employ a single mother wavelet f xed for all textures. In these schemes only the wavelet packet tree of the optimum subband decomposition, and the statistical properties, such as, mean and variance, associated with the marginal densities of its wavelet subband coeff cients are used as features of the texture model. However, [6, 7] noted that the optimal tree for a given texture, and classif cation performance, both vary according to the mother wavelet used in the texture model. Previous studies have also shown that the length, number of vanishing moments, regularity, orthogonality, and degree of the impulse response shift variance of the mother wavelet contribute to the results of wavelet based applications such as image coding [14] and texture analysis [15]. Therefore, an optimal choice of a mother wavelet for the representation of a given texture can be used in feature extraction for a given texture. With a similar point of view, in f lters based methods, where a f lter and a response energy measure is used as the feature extractor, optimization of f lters has been used in order to achieve optimized energy separation [16, 17, 18].
En savoir plus

29 En savoir plus

Texture de la page numérique

Texture de la page numérique

Réfléchir sur les métamorphoses du livre en tant qu’objet conduit à interroger, selon les termes de Mc Luhan, « la spécificité du message livresque 12 » dont l’analyse reste indispensable à qui souhaite comprendre la confection d’un livre et les manières dont une société le reçoit. Auteurs, éditeurs, chargés de saisie, compositeurs et imprimeurs représentent les acteurs centraux de sa construction. Au compositeur revient le rôle délicat de manipuler le texte numérique et de rendre visible la pensée qu’il communique. En faisant correspondre le document au format livre, il établit un lien entre l’auteur, le lecteur et le document. Rapport produit par la construction d’une texture numérique impliquant d’opérer un changement de matérialité duquel est censée naître une meilleure intelligibilité. Dans cette perspective Jean-Henri Martin déclare : « l’écriture n’est plus seulement porteuse d’un discours étant celui des mots prononcés mais aussi celui de la vision », remarque corroborée par Christian Jacob 13 , selon qui l’étude de la composition d’un texte et de sa texture montre « la part déterminante des choix de typographie et de disposition des caractères dans la production du sens, entre les intentions de l’auteur et la réceptivité du lecteur habitué à déchiffrer des codes. Chaque changement dans les techniques de reproduction des textes implique une renégociation globale de ce pacte de lecture avec une part d’oubli des formes éditoriales antérieures ».
En savoir plus

11 En savoir plus

Show all 1975 documents...