FR2660459A1 - Method for segmenting images by analysis of textures - Google Patents

Abstract

The segmentation method processes a source image by: - characterisation of the texture by calculation of attribute vectors with I components associated with several values of analysis resolution; - selection of a discriminating attribute vector with J components by measuring the non-correlation of the texture attributes, by using the modelling of sample zones of K textures, performed in parallel with the characterisation in a processing phase 1'; - classification of the image into K texture classes, for each analysis resolution, based on the modelling of the sample zones and by calculating a confidence criterion associated with the classifications performed; - multi-resolution merging of the partitions obtained on the basis of an analysis using the confidence criterion. On completion of this step, the source image is segmented into K regions. The invention applies, in particular, to the segmentation of aerial images, visible or infrared.

Description

Method for segmenting images by
texture analysis
The invention relates to the field of image processing, and more particularly to a method of segmenting images by analyzing textures.

 The general problem to be solved is to develop a method for analyzing textures with a view to segmenting homogeneous regions of images, visible or infrared, for example aerial images.

 The segmentation of images into homogeneous regions is generally carried out thanks to a characterization of the texture information. This information is all the more important, in the fields of image analysis and segmentation, since these are slightly contrasted and do not present frank structures, of the contour type for example.

Many methods have been proposed for the analysis of textures, methods which it is possible to group into three large distinct families.
- methods using spatial characteristics,
- methods using frequency characteristics,
- structuralist methods.

Among these sets of texture characterization methods, we can cite those which seem to give the most satisfactory results according to the type of images analyzed.
- local luminance histograms,
- curvilinear integration of the luminance signal,
- search for local extrema.

 There are therefore a relatively large number of texture characterization techniques. These methods have often been developed according to the very nature of the texture analyzed and for a given type of approach.

Two types of problems then appear
The first relates to the adequacy of the processing algorithm for the texture studied and the second to the choice of analysis resolution.

 Regarding the adequacy of the algorithm to the texture studied and therefore the choice of characterization parameters, the segmentation of different images, from each of the above methods, shows that, depending on the type of images analyzed, the best result is achieved by different methods. This means that there is a dependence between the type of texture encountered and the most efficient algorithm. This problem is a serious handicap in the development of an automatic image segmentation system by texture analysis.

 Regarding the resolution of the analysis, in most of the texture characterization algorithms, it is necessary to adjust a certain number of operating parameters (thresholds, neighborhood dimension, etc.). In particular, the size of the analysis window is decisive as regards the quality of the segmentation obtained. We observe that there is a dependence between the nature of the textures encountered and the size of the analysis window measuring the textural activity of the images.

 The subject of the invention is a method of image segmentation by texture analysis in which the image segmentation is carried out using a characterization and classification process of the texture, independent of the nature of the image and which does not require iterative adjustments.

According to the invention, a method of segmenting images by analyzing textures is characterized in that it comprises
a texture characterization step in which, for each image point, a set of parameters characterizing the texture and forming a vector of attributes is calculated, this calculation being repeated for several analysis windows of different pillowcases associated with different analysis resolution values,
a step of selecting a discriminating vector resulting from the attribute vector, but of smaller dimension, for each analysis resolution and for each point,
- a step of classification of the image into zones associated with the different expected textures, for each analysis resolution,
- And a step of merging the classification results obtained in the previous step at the end of which the image is definitively segmented between the different expected textures, this fusion being carried out on the basis of a confidence criterion assigned to each point for each of the classifications selected according to the different analysis resolutions.

 The invention will be better understood and other characteristics will appear from the following description with reference to the appended figures.

- Figure 1 is a general flowchart of the segmentation method according to the invention
- Figure 2 illustrates remarkable histograms illustrating the parameters used in the local histogram method
- Figures 3a, 3b and 3c frustrate the parameters used in the curvilinear integration method
- Figure 4 illustrates the spiral integration method
- Figures 5a and 5b illustrate two examples of the distribution of point clouds classified into five texture classes, with respect to two discriminating axes
- Figure 6 represents the Gaussian distributions associated with three different classes of textures
- Figure 7 illustrates the multi-resolution merging step.

In general, the segmentation method according to the invention comprises four processing steps, from a source image, as illustrated in FIG. 1
1. Characterization of the texture by calculating attribute vectors with I components associated with several analysis resolution values, that is to say on the basis of analysis windows of different dimensions;
2. Selection of a vector of attributes discriminating with J components by measuring the decorrelation of the texture attributes, using the modeling of the sample areas of
K textures performed in parallel with the characterization in a processing phase 1 '
3.Classification, i.e. partitioning the image into K texture classes, for each analysis resolution, based on the modeling of the sample areas and by calculating a confidence criterion associated with the classifications performed
4. Multi-resolution fusion of the partitions obtained at each resolution level from a contextual analysis using the confidence criterion. At the end of this step, the source image is segmented into K regions.

 The signal of an image is defined at point M of spatial coordinates M (x, y) by the relation: z = f (x, y). This function z is generally the luminance value at the current point M. The texture at this point has properties related to the variations of z in its neighborhood V (M). The characterization of the texture at a given point must therefore integrate the spatial distribution of the gray levels surrounding this point, in a window defining the neighborhood V (M).

 The image to be analyzed is digitized, that is to say that the luminance of the points which compose it is digitized by quantification on a scale with n gray levels.

 The first three steps of the method according to the invention are carried out for three different analysis resolutions associated with windows of different dimensions.

In the texture attributes calculation phase, two new characterizations, additional compared to the characterizations by more classic attributes also used by the method, were introduced
- The first results from a so-called "spiral" integration
- The second uses the local energies of the gradients.

 In one embodiment, based on the state of the art in texture analysis, seven methods which seem to give the best characterization results have been retained.

 The different texture characterization parameters chosen resulting from this selection of methods are the following I = 14 parameters - module and phase of the local histograms: [MOD, PHA] - curvilinear integral according to 4 orientations: [AO, A45, AOO, A [ 135i, the numbers corresponding to the angles, with respect to a reference, of directions 0. according to which the integrals are calculated - local maximum: [MAX]; - mean and variance: CMOY, VARj - spiral integral: [SPI, MOYSPI] - local energies of the gradients and variance of this energy of the gradients: [NRJAMP, NRJORI, VNRJ].

The MOD module and the PHA phase of local histograms - as defined by G. LOWITZ in a communication entitled "Extraction of caricatures and textures from a local histogram" at the 8th conference on signal processing,
GRETSI, Nice 1981 - are extracted as follows, from local histograms of the luminances established on the blocks of points of a window centered on the current point M.

This method processes the local spatial information of an image.

où n. est le nombre de points de la fenêtre dont le niveau de luminance est le ie niveau de quantification
La phase d'un histogramme local, PHA, est définie comme étant l'indice; le plus petit de l'histogramme maximal hjM le plus proche (au sens de la distance définie ci-dessus D) de l'histogramme local considéré.We show that the distributions of the components of the histograms follow a Poissonnian statistical model. This model introduces an entropy, then a metric distance between these histograms, which makes it possible to define a module and a phase for each local histogram. We then define the two components of the characterization vector as follows
- the module of a local histogram, MOD, is defined as being the distance between the "centered" histogram ho, that is to say the histogram which corresponds to an equal number of occurrences for all the levels of luminance in the analyzed window, such as ho = [1 / n, 1 / n, ...., 1 / nj and the histogram considered h. The distance is then defined as the difference in entropy between these two histograms, so the module is equal to

where n. is the number of points in the window whose luminance level is the ie quantization level
The phase of a local histogram, PHA, is defined as the index; the smallest of the maximum histogram hjM closest (in the sense of the distance defined above D) to the local histogram considered.

 The maximum histogram is defined by M = [O, O, ... O, 1, O, 0, ..., O, O).

hj
j is therefore the index of the maximum component of the local histogram.

 PHA = j (in unit 211 / n).

 By choosing the maximum component, we choose the "dominant" luminance of the local histogram of the analysis window.

 Figure 2 shows the remarkable histograms ho and hi M and a local histogram h.

The curvilinear integration method, for its part - exposed by D. DARBA, J. RONSIN and S. RABOISSON at SPIE
Symposium, Cannes 1985, in a communication entitled "Comparison between co-occurence matrices, local histograms and curvilinear integration for texture characterization -, gives precise information on the distribution of gray levels along lines of analysis of different orientations, ie along privileged directions ei in a non-isotropic texture Figures 3a, 3b and 3c illustrate the parameters used in the curvilinear integration method.

 A measurement of the properties of the texture is obtained by scanning the neighborhood V (M) along lines Li coming from M, according to orientations e. defined above and shown for example in Figure 3a.

 Let x 'be the coordinate of a point of Li in a coordinate system centered in M and of axis Li. A Euclidean metric ds is defined by introducing a scale factor Y between the spatial variations dx' along L. and the corresponding luminance variations dz: ds2 = t2 dx'2 + dz2.

Using the length of the differential arc, we can calculate the curvilinear integral of ds, along the curve
Ci: z = f (x '). For a positive fixed threshold ij, the point Di on Ci is obtained, such that the curvilinear integral of ds, from M to Di, reaches this value y, as illustrated by FIG. 3b.

 The abscissa of D. on L. is ai.

iii
From point M for the different orientations (3i we obtain a set of abscissas [a. Characterizing the texture for this point and forming a characteristic vector which integrates well the spatial distribution of the gray levels in the vicinity of the point and s' therefore adapts to a local measurement of its texture.

 The characterization can be improved by integrating, for each orientation, along a set of parallel lines passing through the neighboring points of M in the direction orthogonal to Li, each texture component then corresponding to an average of characteristics following the same orientation Li as shown in Figure 3c.

 This modification has the advantage of not increasing the dimension of the characterization vector and requires only a simple averaging of quantities already calculated for the neighboring points.

Generally, the characterized orientations correspond to 00, 450, 900 and 1350. The curvilinear integration is carried out with a threshold y of approximately 200 and a weighting coefficient g = 4, corresponding respectively to the stop value of the integration and to a scale factor between the spatial variations dx 'and the luminance variations dz, used to define the Euclidean metric ds for the calculation of the integral as indicated above. The characteristic vector of the current point is then
laO, a45, a90, a135].

The characterization of local maxima, [MAX] - described in particular by O. MITCHELL and W. BOYNE in IEEE
Transactions on Computers, Vol. C25-1977 in an article entitled "A max-min measure for image texture analysisn uses descriptors whose elements are the frequencies of obtaining local extrema in a vicinity of the zone analyzed, under certain conditions of foliage and prior filtering in order to eliminate extrema due to noise.

 This method allows a good characterization of the contrast, the grain and the preferred directions, as well as the periodicities (if they exist) of a texture.

 In a dimension analysis window (33 x 33), the neighborhood (3 x 3) surrounding it is examined for any point in this window called n statistical "1000 points", and it is retained if its luminance is effectively maximum in this neighborhood.

The classical statistical parameters, mean [MOY] and variance [VARl, are calculated for each current point M in a window (nxn) centered on the current point - average: MOY = #M = 1 / N

The characterization by spiral integration is a new variant of the characterization by curvilinear integration of the luminance signal, developed following the following observation
The method of curvilinear integration of the luminance signal, along oriented analysis lines has the disadvantage of being sensitive to a rotation of an angle less than the sampling pitch of the analysis orientations (45 in the example above).

 To eliminate this defect, the number of analysis lines can be increased by increasing the number of orientations.

Unfortunately, this results in a prohibitive calculation cost.

 Hence a new method which consists in calculating the curvilinear integral along a "spiral" centered on the current point. We go from a characterization vector of dimension k (k = 4 in the example) to a vector with a single dimension (a single attribute), and, this measurement is completed by calculation along this spiral, of the value average grayscale.

 Figure 4 illustrates the points of the "spiral" considered for the spiral integration around the current point M.

For each point of the image, two parameters are thus obtained
- a distance characterizing the texture "SPI", for which the integral of ds along the spiral reaches a predefined threshold
- an average value characterizing the component "SPI", "MOYSPI".

Finally, the last three parameters used are calculated by a new characterization method called local gradient energies, not used until now. This method has the advantage of using the information contained in the gradients calculated by a contour detection algorithm called "NEVATIA", - described by R. NEVATIA in an article entitled "Linear feature extraction and description" published in the Flight. 13 of 1980 by Computer Graphics and Image
Processing -.

The convolution of the original image by the masks of "NEVATIA" provides, for any point, two parameters
- the amplitude of the luminance gradient
- the orientation of the luminance gradient.

 The texture characterization method then consists in extracting information from the two aforementioned parameters, in an analysis window centered on the current point.

 First, an image of the orientation "activity" is calculated by simply detecting "contours" on the orientation image.

 Secondly, for each point of the image, the energy, E, of the images of the gradient amplitudes and of the orientation activity is calculated in a window (nxn) centered on the current point.

The energy E, in a window (nxn) for a point M (x, y) y) is given by

 with z (x, y): luminance signal at point M (x, y).

 We thus obtain, at all points, the two parameters amplitude and orientation of the luminance gradient NRJAMP> NRJORI and the energy variance E, VNRJ is calculated from the parameter NRJAMP.

 Thus the I = 14 characterization parameters X. chosen are calculated for each image. Most of them are calculated in a window centered on the current point, sliding on the image.

 But, as indicated above, in most of the texture characterization algorithms, it is necessary to adjust a certain number of operating parameters: (thresholds, dimension of the neighborhood, ...). In particular, the size of the analysis window is preponderant, because there is a dependence between the type of texture present (resolution) and the size of the analysis window.

As indicated above, the texture characterization and segmentation method according to the invention is independent of the nature of the image and does not require iterative adjustments. For this, as specified above, the size of the analysis window is not fixed a priori nor possibly modified iteratively:
In its first step 1 in FIG. 1, the method consists in calculating all of the characterizing parameters, or attributes, mentioned above, already used in the various known methods or resulting from new characterizations, systematically for each of three sizes of analysis windows, corresponding to different analysis resolutions
- window (33x33) pixels, corresponding to the statistical minimum of maximum useful dimension for the images to be processed
- window (7x7) pixels corresponding to the minimum size of perception in the images
- window (15x15) pixels which is a window of intermediate size compared to the previous two.

The second step, step 2 in FIG. 1, is the selection of a "best" vector of characteristics or discriminating vector from all of the parameters measured, and this for each of the three predetermined analysis resolutions corresponding to the windows d described above
For this selection of the discriminating vector, the method of selecting the texture parameters implements a process of searching for the best linear combinations of these parameters Xi, by a discriminating analysis, to replace the
I initial variables by J new variables, with J <I.

 The advantage of such an approach lies in the fact that a reduction in the number I of attributes is thus effected (14 attributes for each point initially in the example used), by minimizing the loss of information, which allows a gain of very significant time when classifying into zones, even if the number of initial variables is limited, because as soon as the size of the samples is large, which is generally the case in image processing, the number of values to be processed can to be limited in an interesting way.

This method consists in determining an orthogonal transformation which allows the projection of the point cloud of the I dimensional hyperspace into a dimension subspace.
J as weak as possible, and this with a minimum of deformation and therefore of information loss, by the discriminant factor analysis method
For this, we use knowledge on the number of classes and on learning sample areas belonging to these classes, acquired in phase 1 'of modeling areas of different textures, as follows
Let K classes be recognized, and a partition of N training samples into K subsets of respective sizes ENi 'N2, ..., Nk, ..., NKJ.

The k subset corresponds to the set of samples of class C k in the training set. 1l is given by: Xk = # Xkn1 with nt [1, Nkj
His average is worth

Its covariance matrix is

The overall average calculated over the N training samples is

The covariance matrix over the set of N training samples is

et B est la matrice de covariance inter-classe

The calculation shows that there is a relationship between the covariance matrix of the set, T, and the covariance matrices Tk of the subsets, which is written T = W + B,
where W is the intra-class covariance matrix

and B is the inter-class covariance matrix

 The total covariance matrix T of the point cloud formed by K classes is equal to the sum W of the covariance matrices of each of the K classes and of the covariance matrix B that this cloud would have if the "mass" of each of the classes was concentrated at the "center of gravity" of the class, ie if all the samples in the class had the same mean value.

 The goal of discriminant factor analysis is to obtain a representation of the cloud in a J-dimensional subspace, so that the class projections are the best separated and each of them is the most homogeneous.

 This can be expressed using the intra-class and inter-class variances extracted from the covariance matrices W and B.

The intra-class variance is defined by

The inter-class variance is given by

with the relation: var (XT) = var (Xw) + var (XB)
The analysis of the var (XB) / var (Xw) ratio is interesting.

 If this ratio is high, it means that the classes form homogeneous sets well separated from each other.

 It will be easy to separate the classes.

 On the other hand, if it is low, the separation of classes will be more difficult.

Good class discrimination can be obtained if
- the intra-class variance is minimum and,
- the inter-class variance is maximum.

 The discriminant factorial analysis then searches for the subspace of dimension J, in which the ratio var (Xb) / var (Xw) is maximum.

 The ratio of the inter-class variance to the intra-class variance of the projected cloud is maximum when the direction of the projection axis is defined by the eigenvector associated with the largest eigenvalue of the matrix (W 1) .B ).

 The discriminant factor analysis performs the projection of the point cloud on the subspace generated by the J eigenvectors associated with the J largest eigenvalues of the matrix (W # 1.B).

 We can also note, that the inter-class covariance matrix B is at most of rank K '= (K-1), if we have K classes.

If W is invertible, the matrix (W 1.B) is also of rank
K '. One thus obtains only K 'nonzero eigenvalues. It is therefore useless to project the point cloud in a space of dimension higher than the number of classes K minus one.

We can just as well try to maximize the ratio
UtBu / utTu which is an increasing function of Ut. Bu / Ut .Wu when W is not invertible. In this case, one must seek the J principal eigenvectors of the matrix (TB).

This discriminating factorial analysis (AF D.) Therefore makes it possible to obtain in the representation of the main inertia planes, projections of the classes which are, on the one hand, the best separated from each other, and on the other hand, the most homogeneous, taken one by one. In general, this is expressed using the following criterion: the inter-class variance must be maximum and the intra-class variance minimum and stefect as follows
From W, intra-class covariance matrix, and from
B, inter-class covariance matrix, the eigenvector [a] such that the ratio atBa / atWa is maximum is sought.

 For this, the product matrix (W B) is diagonalized. The eigenvalues (el, e2, ..., ek, ... em) ... em) are ranked in descending order, and the associated eigenvectors (al, a2, ..., ak, ..., am) correspond to the different factor axes.

 In a system of discriminating axes Yi-Yj, we can represent the clouds of projected points each corresponding to a texture class. We generally choose as the representation plane that of the first two discriminating factor axes Y1 and Y2, as shown in Figures Sa and 5b.

coordonnée sur le ler axe discriminant

coordonnée sur le 2ème axe discriminant i : indice de la classe j : indice du point.A point has coordinates in this new coordinate system

coordinated on the 1st discriminating axis

coordinate on the 2nd discriminating axis i: class index j: point index.

ceci en centrant les axes du centre de gravité du nuage de points d'une classe.The center of gravity of the points of a class has for coordinate on the 1st discriminating axis

this by centering the axes of the center of gravity of the point cloud of a class.

Two cases of discrimination for 5 classes of texture have been shown in Figures Sa and 5b.

 In the case presented in FIG. 5a, the discrimination is easily carried out due to the good separability of the classes and the homogeneous character of the point clouds. Conversely, the case of FIG. 5b is more difficult. Only classes 5 and 2 seem to give homogeneous clouds, discrimination will be less easy for the other classes, especially since the inter-class variances are extremely small.

 Following this modeling intended to select the discriminating axes (linear combinations of characteristic parameters) from the different samples of the distinct textures identified in the image, the classification of the image is carried out.

 The purpose of the classification is to obtain the segmentation of the image into several classes thus designated by the samples identified by the operator and which have been modeled. The number of classes, linked to the number of homogeneous zones is known a priori and has been used to learn the decision system; this system then assigns, in the classification phase, all of the points to one or other of the classes, in a so-called supervised classification.

 For this, the learning windows also make it possible to establish the normal Gaussian laws, used for obtaining the discriminating functions: a good classification must find a maximum number of points assigned to the corresponding class C. in the associated learning window .

 In order to carry out the selection of the discriminant vector and then in the next phase the multi-resolution fusion of the segmentations obtained for the different analysis resolutions, a local measure of the quality of the classification, or segmentation, using a confidence criterion is calculated during of the classification phase. The quality of the resulting segmentation can thus be improved. Multi-resolution fusion takes into account the fact that a large analysis window is interesting for the characterization of homogeneous areas of large surface, while on the other hand, a small analysis window will give better results in textured areas of reduced size.

For this purpose, the segmentation process is established with a multi-size selection for the three dimensions of the analysis window described above associated with three different resolutions:
R1: (7x7), R2 (15x15), R3 (33x33).

 An analysis of the neighborhood of the current point makes it possible to retain, among the classes learned during the modeling phase 1 ′, the most probable class, and to assign to this classification a resulting rate of "confidence".

In addition, a confidence rate measure is established on the discriminating functions to estimate the reliability of the classification.
Let be a Gaussian distribution of order N, where N is the number of parameters combined, the likelihood P of the assignment of a point M of the image of parameters X = # X1 ... Xi ...

XI) to a class Ci is defined by
P (X / Ci) = exp (2-Yi / 2) / (2 #) N / 2 (det (#i)
with Ei = covariance matrix and Y. normalized data deduced from Xi such that

 where CL is the mean vector whose components lli are the means of the various parameters X ..

either Log P (X / Cp + Log [(2 #) N / 2] = [Log [det (#i)] + Yi]
We pose
H1 = - [Log [det (#i) + Yi]
The study of H1, for each point makes it possible to determine the class of the point. In calculating the confidence rate, we reason on absolute values of H1.

As an example, Figure 6 represents the Gaussian distributions of the likelihood of assignments to three distinct classes, C1, C2 and C3 as a function of X. For a point
X, the likelihood of assignment to C1, C2 and C3 are H1C1, H1 C2 and H1C3 respectively. The classification of the point by the maximum likelihood method is carried out and a point confidence criterion H is calculated.

- For the abscissa point X in Figure 6, H1C1 being the maximum of HiCi, X is classified in C1 and the assignment likelihoods are classified: here # H1C1, H1C2 and H1C3}
- The confidence measure takes into account the difference
Hlmax (here H1C1) and next H1 (here H1C2).

- Then the percentage characterizing the following point confidence criterion H is calculated
H = (next Hlmax-Hl / Hlmax) x 100
- Finally, a weighting of H as a function of the deviation of point X from the average of the chosen class is carried out.

 The synoptic diagram of the multi-resolution fusion processing is illustrated by FIG. 7. At this stage of the method three segmentations of the image corresponding to the three analysis resolutions R1, R2, R3 are known as well as the confidence images associated with these segmentations. FIG. 7 symbolizes the segmentation of the image into two zones A and B for the three resolutions and the associated confidence images CA, C B.

 The multi-resolution fusion is then based on the analysis of the neighborhood of the current point, in the segmented images.

This contextual analysis takes into account the confidence rates calculated for each of the three sizes of analysis windows, and performs a majority sort on this neighborhood.

 For the fusion, the class retained Ci (or A or B on figure 7) is that for which the number of points assigned to the class k is maximum, in the neighborhood considered and the confidence of the current point M corresponds to the average confidence points of this class.

 This method gives greater detail (on the borders of the different textures in particular), and greater homogeneity of the areas of large surface textures. The main result obtained by this process is therefore that it allows to synthesize an optimal characterization of the texture, from the best state of the art models.

 Despite the disadvantage of a certain over-cost of computation compared to the application of an algorithm implementing a unique modeling of the texture, the method converges quickly towards a concise characterization of the texture and a quality segmentation. In addition, the disadvantage of additional cost is not prohibitive for an interactive exploitation, because the architecture of the process is easily parallelizable.

 On the other hand, in the vast field of aerial imagery applications, the implemented process is of considerable interest, because of its particular qualities, in particular its autonomy, a gain in terms of the confidence attached to the characterization of texture and its possibility. openness to other characterization methods.

Almost all of the texture analysis methods known from the state of the art, allow good quality segmentation, but often with heavy operating constraints.
- either the model used is appropriate for a very particular category of images of homogeneous textures,
- either the adjustment of the control parameters must be refined iteratively on each type of image,
- either the number of texture characterization attributes is too high.

 The method according to the invention is autonomous in the sense that it does not require any control parameter, and that it automatically adapts to the image analyzed. This quality becomes paramount when we observe that in aerial imagery, very few textures are homogeneous over large areas.

 To evaluate the result of an image segmentation, it is moreover necessary to have a fine measurement of the power of the link which attaches each pixel of the source image to a texture class. This point measure is defined in the process by the degree of confidence, and the confidence image thus generated is an important criterion of the quality of segmentation.

 In addition, this measurement can be crucial for applications requiring automatic interpretation, since it weighs the reliability of the information extracted.

 Thanks to its contextual analysis method, the developed method produces segmentations whose degree of confidence is better than those obtained in segmentations by state-of-the-art algorithms.

 In addition, given the current efforts made on texture analysis, it is likely that in the future, new texture characterization models will be created. The structure of the process is by definition open to the integration of such models, and also makes it possible to assess their contribution compared to existing models. This facility is particularly advantageous for improving, over time, the quality of the segmentations produced.

Claims (6)

 1. Method for segmenting images by analyzing textures, characterized in that it comprises  - a texture characterization step (1) in which, for each image point, a set of parameters characterizing the texture and forming a vector of attributes is calculated, this calculation being repeated for several analysis windows of different sizes associated with different analysis resolution values,  a step (2) of selecting a discriminating vector resulting from the attribute vector, but of smaller dimension, for each analysis resolution and for each point,  a step (3) of classifying the image into zones associated with the different expected textures, for each analysis resolution,  - And a step (4) of merging the classification results obtained in the previous step at the end of which the image is definitively segmented between the different expected textures, this fusion being carried out on the basis of an assigned confidence criterion at each point for each of the classifications selected according to the different analysis resolutions.  2. Method according to claim 1, characterized in that the discriminant vector associated with each point, for each analysis resolution, has components resulting from linear combinations of the attributes of the attribute vector, these combinations being determined so that the variance between the samples of the same texture is minimal and that the variance between samples of different textures is maximum from a modeling of samples of textures selected by the operator in the image.  3. Method according to claim 2, characterized in that the classification of the image carried out for each of the analysis resolutions is based on the distribution of the components associated with the different texture samples selected by the operator and on the values of the components corresponding to the discriminating vector associated with the point to be classified with respect to these distributions.  4. Method according to one of claims 1 to 3, characterized in that the attributes calculated for each point come from several methods of texture analysis, so that any texture analyzed is recognized by at least one of the parameters characterization.  5. Method according to claim 4, characterized in that the characterization parameters calculated from the luminance values of the points in the vicinity of the current point are  - a local histogram module and phase  - curvilinear integrals calculated according to four different orientations coming from the current point to characterize  - a local maximum  - a mean and a variance  - a spiral integral calculated from points surrounding the current point in a spiral  - the local energy of the gradients and the variance of this energy of the luminance gradients.  6. Method according to one of the preceding claims, characterized in that three different resolutions are used resulting from the analysis corresponding first to windows of 33x33 image points associated with the statistical minimum, the second to windows of 7x7 d points 'images, corresponding to the minimum size of perception in the images, and the third to windows of 15x15 points of images, of intermediate size between the two previous ones.