WO1999003065A1

WO1999003065A1 - Method for detecting and/or determining characteristics related to remarkable points of an image

Info

Publication number: WO1999003065A1
Application number: PCT/FR1998/001418
Authority: WO
Inventors: Naamen Keskes; Pierre Baylou; Sébastien Guillon
Original assignee: Elf Exploration Production
Priority date: 1997-07-07
Filing date: 1998-07-02
Publication date: 1999-01-21
Also published as: EP0923764A1; NO990883L; FR2765707B1; FR2765707A1; OA10989A; CA2264903A1; NO990883D0

Abstract

The invention concerns a method for detecting and/or determining characteristics related to remarkable points in an image, characterised in that it consists in evaluating the local dip variability of a first point in the image relative to at least another point located in the proximity of said first point, by computing the local anisotropy on said point gradient field, said anisotropy being dependent on terms related to a dispersion of orientations and to the module of gradient vectors, and at least one of said terms is weighted.

Description

Method for detecting and / or determining characteristics linked to remarkable points in an image

The present invention relates to a method for detecting and / or determining characteristics linked to remarkable points of a given multidimensional image and capable of being implemented, in particular for the detection of virtual contours which may be present in said image.

The virtual contours result from the proximity of remarkable points grouped or not grouped in elongated supports, said remarkable points being in particular junction points, angular points or even termination points. A human observer, when studying or interpreting an image, whatever the nature of said image, attempts to determine the contours of what could be considered to be a characteristic event of the image. The transposition of this faculty of observation of the human eye into an automatic interpretation system proves to be very delicate and it is, in general, difficult to develop and apply operators giving an account of a subjective impression. In general, a virtual contour, sometimes designated by the expressions of subjective or perceptual contour, is identified by the eye in two stages, one detecting the endings and the other trying to establish the relations of connectedness between said endings. There are methods capable of detecting remarkable points of an image. One of the methods is due to the work of J.P. COCQUEREZ and S. PHILIPP described in a book entitled "Image analysis: filtering and segmentation" (MASSON, 1995). Another method, due to the work of P. BAYLOU et al. , is described in an article entitled "Evaluation of the anisotropy of textures. Comparison of the methods applied to the characterization of materials" GRETSI 95, JUAN LES PINS, pages 1245-1248, 1995.

However, these methods propose to calculate the global anisotropy of an image which is given by the formula: G, G, i '.. ⁾

Iso = kl

G, G, d) ι.J

in which

VS\ ; are pairs of neighboring pixels (in the sense of 4v),

G _j and G; are the modules of the gradients at positions i and j of two neighboring pixels, and cq; is the angle formed by the gradient vectors of pixels i and j.

Such determinations of the overall anisotropy make no difference between the strong and weak gradients and therefore have a weak separating power or a weak discriminating power since the gradient vectors of all the pixels are taken into account. In addition, they require to compute all the angles formed between the gradient vectors as well as their arguments because cq j = arg (Cq) -arg (Gi). Finally, these methods are very sensitive to noise and do not allow a distinction to be made between gradient vectors attached to a remarkable point and those attached to noise. Thus, the selectivity obtained by these methods is really very reduced.

The object of the present invention is to propose a method which makes it possible to transform a given initial image into another image or representation in which specific characteristics of the initial image are highlighted. This method considers the directional coherence of the gradient vectors according to cliques of order 2.

Another object of the present invention is to propose a method which is applicable whatever the nature of the initial image, such as for example a seismic image, a medical image in which one wishes to highlight in particular the branching or the division blood vessels, an aerial image to highlight in particular a crossing of roads, a place or a characteristic object. In what follows we give the example of a seismic image in which we wish to highlight seismic characteristics relating in particular to seismic horizons, to zone boundaries (contours), to chaotic or channeling textures, without considering this enumeration as exhaustive. An object of the present invention is a method of detecting and determining characteristics of a multidimensional image linked to remarkable points of said image, which is characterized in that it consists in evaluating the variability of a local dip of a first point of the image with respect to at least one other point located in the vicinity of said first point, by calculating the local anisotropy on the field of gradients of said point, said anisotropy being dependent on terms related to a dispersion of the orientations and to the modulus of the gradient vectors, and in that at least one of said terms is weighted. According to another characteristic, each of said terms is weighted.

According to another characteristic, the weighting is carried out by raising at least one of said terms to a power.

According to another characteristic, each of the terms is raised to a power whose power factor is different from one term to another.

According to another characteristic, the weighting is carried out according to the nature of the remarkable point to be detected.

According to another characteristic, the method is applied to a seismic image. According to another characteristic, the points to be detected are points selected from the junction points, the angular points or the end points.

An advantage of the present invention resides in the fact that only the remarkable points which have a discontinuous character are detected, and this, by means of an operator for measuring the local anisotropy of the field of gradients around said remarkable points because there is noted that locally, along a contour, the gradients have a strongly anisotropic distribution with a dominant direction orthogonal to said contour, while at the remarkable points the distribution of the gradients is more isotropic.

Another advantage of the present invention is that the operator is configurable, which makes it possible to adapt it as a function of the type of remarkable point to be detected and / or to be highlighted as well as the signal / noise ratio. In other words, we can adjust the weighting of each operator terms depending on the nature of the remarkable point to be detected.

Other advantages and characteristics will emerge on reading the description of the method according to the invention, as well as the appended drawings in which:

FIG. 1 is an image extracted from a seismic section having a fault and a channel,

- Figure 2 is a representation of the anisotropy of the image of Figure 1, - Figure 3 is an image extracted from a seismic section having a channel,

- Figures 4 and 5 are representations of the anisotropy of the image of Figure 3.

The present invention uses an operator which makes it possible to analyze the local dispersion of the orientations of the gradient vector of the pixels of an image. Indeed, locally and along delimiting contours, for example a fault or a channel, like those represented in FIGS. 1 and 3, the gradients of the points (pixels) have a strongly anisotropic distribution, that is to say that the dominant direction is orthogonal to the contour, except at break points, angular points or triple points on which the distribution of the gradients is more isotropic. According to the invention, the degree of anisotropy of the distribution of the gradients is measured.

In order to evaluate the variability of the orientation of the gradients, the anisotropy is determined locally for each pixel of an image rather than globally over the entire image. The determination which is carried out consists in calculating the orientation differences for pairs of neighboring points, for example for pairs of neighboring pixels, said differences being weighted by the modules of the gradients of the pixels considered.

If we used the general formula (1), we would have to measure the arguments of the gradients since the angle oq between the two directions is equal to arg (G *) - arg (G ^.

To avoid measuring angles α er the numerator of formula (1) by ∑G, G y, s ^~ m

gives the following formula:

The calculation which results from the application of formula (2) is less costly since only multiplications are to be carried out.

The determination of the anisotropy carried out using this calculation proved to be insensitive to the amplitudes of the gradients, the zones of weak gradients being able to give results similar to those obtained for zones of strong gradients.

Another aspect of the invention is to change the normalization factor which is represented by the denominator of formulas (1) or (2), which leads to favor the pixels of strong gradient. The determination of the anisotropy to be carried out is then calculated by the formula:

The local anisotropy calculated with formula (3) then makes it possible to detect only the characteristic points of strong gradient whereas with formula (2), the points are detected independently of the norm of their gradient, only the angular dispersion of the gradients.

The points to be detected having different properties depending on the type of image to be processed, it is proposed according to another aspect of the invention to weight the angular dispersion factor of the gradients represented by the numerator of the formula (3) and / or the normalization factor.

In this way, the anisotropy is measured using the formula:

in which

Gi and G; are the modules of the gradient vectors G \ and Gj with the neighboring pixels i and j considered and which belong to the set of cliques of order 2, in the sense of the neighborhood 4v, in the plane the gradient used being in particular the gradient of DERICHE as explained for example in the book by JP COCQUEREZ and S. PHILIPP "Image analysis: filtering and segmentation", MASSON, 1995, n is a parameter which manages the influence of the amplitude of the gradients, - p is a parameter which manages the influence of the angular variation of the pairs of vector gradients on the anisotropy, q is a parameter which also manages the influence of the amplitude of the gradients and which must be combined with the parameter n, because when one increases the parameter q, one limits the effect of the parameter n and one attenuates the contribution of strong gradients.

When it is desired to detect the points whose gradient is of high amplitude, constituted for example by well-marked contour points, then the parameter n is increased and the parameter q is decreased. This makes it possible to limit false detections in noisy regions in which the points have an isotropic gradient field but small amplitudes. When we increase the parameter p, the sinP function becomes very selective around π / 2 and we only detect at the limit the points of the contours located on a right angle. Thus, when the parameter p is increased, the large angles are favored and the anisotropy is then sensitive only to strong angular variations.

From the foregoing, it can be seen that, depending on the nature of the points which it is desired to detect, it is possible to give each of the three parameters n, p and q the values most suitable for these points.

Examples of application are given with reference to FIGS. 1 to 5.

FIG. 1 is an image extracted from a seismic section presenting a fault F and a channel C. In order to detect the zones of ruptures which include terminations or triple points and which are associated with regions of strong gradients, we take n = l, p = 1, 5 and q = 0.5. The choice of n greater than q (n> q) makes it possible to take more account of the contrasting regions. The choice of p greater than 1 makes it possible to be more selective on the dip variations.

The calculation of the anisotropy according to the method of the present invention for the image leads to a representation of said anisotropy as as shown in Figure 2. We see in Figure 2 that the fault F and the channel C are well detected.

For the image of Figure 3 which is extracted from a seismic section, we want to highlight the apparent presence of a channel. For this image of FIG. 3, two determinations of the anisotropy were carried out, the results of which are shown in FIGS. 4 and 5.

The first determination is made with: n = 1.2; p = 0.5; q = 0.5 and gives the envelope of the channel which is well detected as it appears on figure 4. The second determination is carried out with: n = 2; p = 2; q = 0.5, which allows remarkable points to be obtained inside the channel, as shown in Figure 5.

According to another aspect of the present invention, it is also possible to carry out a filtering. For this, the anisotropy is calculated locally on an observation window, by summing the terms on the pairs of points which are neighbors in the 4v sense. A 5x5 or 7x7 size window is enough to give good results. If we want to weight the contribution of each pair of points as a function of their difference in the center of the observation window, we can use a low-pass filter of the DERICHE type whose impulse response is:

The choice of allows you to manage the width of the calculation window; in practice we choose α ≈ 1.

Two images are then calculated, one being representative of the numerator and the other being representative of the denominator of formula (5), considering only the pairs of neighboring points including the current point.

Each component is then filtered recursively before calculating the isotropy as the ratio of two filtered images.

The transition from a two-dimensional (2D) image to a three-dimensional (3D) image is a simple extension by taking for each point the 3D gradient and considering the neighborhood in the sense of 6v.

Claims

1. Method for detecting and determining the characteristics of a multidimensional image linked to remarkable points of said image, characterized in that it consists in evaluating the variability of a local dip of a first point of the image by relative to at least one other point located in the vicinity of said first point, by calculating the local anisotropy on the field of gradients of said point, said anisotropy being dependent on terms related to a dispersion of the orientations and to the modulus of the gradient vectors, and in this that at least one of said terms is weighted.

2. Method according to claim 1, characterized in that each of said terms is weighted.

3. Method according to claim 1 or 2, characterized in that the weighting is carried out by raising at least one of said terms to a power.

4. Method according to claim 1 or 2, characterized in that each of the terms is raised to a power whose power factor is different from one term to another.

5. Method according to one of claims 1 to 4, characterized in that the weighting is carried out according to the nature of the remarkable point to be detected.

6. Method according to one of claims 1 to 4, characterized in that it is applied to a seismic image.

7. Method according to claim 6, characterized in that the characteristic points to be detected are points selected from the junction points, the angular points or the termination points.