FR2783950A1 - Processing of digital image to automatically locate and extract one or more objects in the image - Google Patents

Abstract

The location of an object is in four steps. The image is segmented into homogeneous regions (Rp(S)), which are labelled as one of: not labelled, significant, or non-significant. Neighboring significant regions are fused into zones (Z1 - Z4), and boxes (B1-B4) enclosing zones determined. An intercept vector is used to construct the object contour, which is refined by filtering.

Description

METHOD AND SYSTEM FOR PROCESSING AN IMAGE

  DIGITAL, FOR EXTRACTING DIGITAL OBJECTS

  The present invention relates to a method for automatically locating and / or extracting one or more digital objects in an image that is also digital. It also concerns a system for the implementation

of this process.

  In the context of the present invention, the term "object" should be understood in a general sense. It is a region of a digital image, homogeneous according to certain predetermined criteria which will be specified below, being able to

  different from other regions and / or from the background (back

  plan) of this digital image, whether this background is homogeneous or not, as it will be shown. By way of non-exhaustive example, an "object" can be constituted by an object, in the usual sense of the term, but also by a character, or more generally by any closed contour defining a

  image area of particular features.

  The extraction operation is often called

  "Trimming" in image processing methods.

  The most rudimentary methods consist in determining entirely manually the contour of an object that is to be extracted from an image. To do this, the software for image processing, have various functions, for example the so-called "lasso" function that draws around the object with the aid of a pointing device (for example a "lasso"). mouse ") a contour approaching the contour of the object to be" cut off ". We can then select the surrounding area, or its logical complement, and perform a cutting (function "scissors") and "stick" the image and on a homogeneous background. It is also possible to erase the areas surrounding the object using a function called "eraser", that is to say in reality also apply a uniform color on these areas. The two methods are combinable: a "rough" clipping can be obtained by selecting an area surrounding the object to be cut off, followed by a selective erasure of the remaining parasitic zones, advantageously by magnifying the image ("zoom" function ). The object thus cut can then be used for various purposes, in particular for a superposition on a different background background or even a new image (special effects, etc.). It is then sufficient to color the new background in a color not present in the cut-out object and to associate with this background an attribute called "transparent color". We can then superimpose the new image (object + background said "transparent") with any other background. This technique is the one well known

  "the blue background", commonly used in video.

  There are other more sophisticated techniques for extracting the outlines of an object in an image. we can evoke, for example, the well-known process

  under the general term of "magic wand".

  However, all these techniques are of the manual type and therefore require a very important time to obtain the desired result. On the other hand, they depend on the dexterity of the operator to obtain a suitable result, if not perfect. Finally, they do not allow

  usually only extracting one object at a time.

  Also, it has been proposed more efficient methods, and partially automated, using so-called techniques of "active or flexible contours". These techniques are based on the concept of

  "snake" or "snake" according to the English terminology.

  "Snakes" are defined outlines

  manually by an operator via a "human-

  machine. "They can then self-adapt to more accurately represent the actual contours of an object that is desired to be trimmed, but they require the initial step of manual trimming, which can be called" coarse " From a mathematical point of view, a "snake" can be associated with a so-called "spline" function of energy minimization, since a "snake" is defined as an energy function. is to minimize this function, in order to find the best adaptation between the "snake" and the real contour of the object to be cut.The "serpents" have the property to modify dynamically, and their shape, and their position, during the search of a state

minimal energy.

  By way of non-exhaustive example, for a

  more detailed description of the process that has just been

  recalled, we can refer profitably to the article by M. KAS et al., entitled "Snakes: Active contour models", published in "International Journal of Computer Vision",

flight. 1, No. 4, pp. 321-331, 1987.

  It is readily apparent that such methods have advantages over conventional object extraction methods. They are especially autonomous and self-adaptive. They can be made robust to image resizing by using a Gaussian smoothing function during

calculating the energy of the image.

  But they continue to have disadvantages

non negligable.

  First of all, as it was recalled, it always involves a manual step, to define the "snake"

initial.

  On the other hand, the accuracy of the contours is determined by the convergence parameter adopted for the minimum energy search step. As the calculation time is proportional to the precision of the contours requested, it is not possible to obtain, in a single iteration, the

  real contours of the object to be cut.

  In addition, the active contour type processes are noise-resistant only in a very relative manner. Indeed, the experiment shows that they do not make it possible to detour an object if the digital image tested is very raster or has defects of digitization (effects of blocks due to the poor quality of the digital video, noise

local but not global, etc.).

  The object of the invention is therefore to provide a method for locating and / or extracting digital objects in an image that is also digital and intended to overcome the disadvantages of the methods of the known art, some of which need to be recalled, as well as that a system for implementing this method. In a preferred embodiment, it proposes a fully automatic method of contour detection, that is to say without external intervention, very precise, then allowing to extract a digital object to the pixel, in the form of polygons to n sides, with n 2 3. The process does not involve the use of convergence, which

  makes contour detection very fast.

  To this end, in addition to a preliminary phase of acquiring a digital image, which is common per se, with most of the methods of the prior art, the method according to the invention comprises, in a preferred embodiment the following phases and steps: a / a first phase of locations of distinct digital objects in said digital image. This phase is broken down into four successive stages: 1.- segmentation of the image into homogeneous regions; 2.The labeling of these homogeneous regions using a particular state attribute, including the following states: a state labeled "Not Labeled", a state called "Not Significant", and a state called "Significant" ; 3.- merging neighboring regions into zones associated with the status attribute called "Significant"; 4.- and the determination of so-called "box" areas encompassing each of said zones, advantageously of rectangular shape; and: b / a second phase consisting in the actual extraction of the distinct digital objects, which itself comprises, for each of the aforementioned "boxes", three successive steps: 1.- the acquisition of a set of vectors of interception with the aforementioned areas; 2.- the construction, from this set of intercepting vectors, of an outline of the object, said primary; and 3.- the reconstruction of the real contour of the object to

  extract from the primary contour.

  The last step consists notably in smoothing the polygon obtained in the previous step, in detecting and eliminating the discontinuities of this polygon, and more generally in suppressing parasitic effects due to a

imperfect digital image.

  It should be understood, however, that in a further embodiment, the method may not include

than the second phase.

  Indeed, if we know, a priori or by examining the image displayed on a screen, that it includes only a significant object, the edges thereof may constitute the sides of the only "box "necessary to achieve the stages of the second phase. These sides form the base from which the vectors

  interception will be determined.

  More generally, the second phase can be initialized simply by manually drawing on the digital image a "box", that is to say a rectangle, encompassing an object to be extracted. Certainly, this particular step is then manual, but it does not require dexterity or particular minutia on the part of the operator. It can be performed in a very short time, for example using a pointing tool, type "mouse", and delineating the two end points of the rectangle describing the "box" selection, on a computer screen displaying the digital image. Then, the subsequent steps, that is to say those belonging to what has been called the second phase, take place fully automatically, while retaining the advantages of the method according to the invention: accuracy of the clipping, speed etc. At the end of the second phase, in the preferred embodiment of the method according to the invention, a user has, without any outside intervention, the list of all the distinct objects present in a given digital image, in the form of polygons. This provision is also a further advantage of the invention. The computer representation of these polygons is available in the form of digital signals that can be stored directly in memories of an information processing system, also fully automatically, and subsequently exploited at a later date.

various purposes.

  In a further variant, the extraction of objects and the storage of the digital signals representing them can be extended to a sequence of video images, and no longer only to static images, with the possible detection of redundancies and storage of the images.

  only single-occurrence objects in the above-mentioned sequence.

  The subject of the invention is therefore a method for processing a digital image in order to extract digital objects constituted by regions of pixels with homogeneous optical characteristics, said regions having a closed contour, characterized in that it comprises a phase of locating at least one of said digital objects comprising at least one step of creating, in said digital image, a closed area encompassing this digital object, the coordinates of the pixels of the component being determined with respect to a reference frame determined related to said closed area; and an automated extraction phase of at least one digital object comprising the following steps: a / acquisition of a set of interception vectors with said contour of the object, each representing a coordinate of a point of this contour with respect to said reference frame, the interception being determined by the prior fixing of a threshold parameter, characterizing said homogeneity and the detection of the variation of this parameter beyond this threshold; b / the acquisition, from this set of interception vectors, coordinates, with respect to said reference, points describing a first contour of the so-called primary digital object; and c / the reconstruction from this so-called primary contour of a second contour representing the real contour of said object to be extracted, by eliminating parasitic noise

and smoothing said primary contour.

  The subject of the invention is also a system for

implementation of this method.

  The invention will be better understood and other features and advantages will become apparent upon reading

  the description which follows with reference to the appended figures,

  among which: - Figures 1A to 1H illustrate the main steps of a first phase of the method, according to a preferred embodiment of the invention, or phase of locating objects contained in a digital image; FIGS. 2A and 2B are diagrams schematically illustrating the neighborhood relations between pixels of a digital image, according to two methods; FIGS. 3A to 4D illustrate the main steps of a second phase of the method, according to a preferred embodiment of the invention, or extraction phase of the objects contained in a digital image; FIGS. 4E to 4I illustrate particular operations executed during the last step of the object extraction phase; FIG. 5 schematically illustrates the architecture of a digital image processing chain comprising a system for locating and / or extracting digital objects contained in a digital image, for implementing the method according to FIG. invention; and FIG. 6 schematically illustrates the main components of an exemplary system for locating and / or extracting digital objects contained in a digital image, for implementing the method according to the invention. The method of locating and / or automatically extracting objects in a digital image according to a preferred embodiment of the invention will now be described in more detail. Indeed, according to this embodiment, the method comprises two phases: a first phase allowing the location of the digital objects in a given digital image and a second

  phase allowing the actual extraction of these objects.

  Each phase has several stages, as it has

been indicated.

  The first phase will be described by reference to

  Figures 1A to 1E, and Figures 2A and 2B.

  We will first define more precisely what is called "digital object" (which will be called hereinafter

  simply "object") in the context of the process of the invention.

  It is a set of homogeneous regions of pixels of a digital image. For each pixel, there is at least one other pixel that is adjacent to it in this region, which is considered as its neighbor in relation to a relation of

  neighborhood, and which can be of one of the types called "4-

  neighborhood "or" 8-neighborhood "for a two-dimensional array Figure 2A shows a so-called" 4-neighborhood "type relationship The central pixel PO is surrounded by four pixels, Pl to p4, arranged crosswise, in two orthogonal directions: Figure 2B shows an "8-neighborhood" type relationship The central pixel p0 is in the center of eight pixels, P1 to P8, which surround it

  completely (the eight pixels forming a square).

  Generally, the digital image is a color image comprising several channels, for example three channels for "RGB" encoding. The method is, however, entirely compatible with a single-color image, especially in grayscale. It then has only one channel,

  which we will call in a generic way of color.

  From there, one can more precisely define a homogeneous region. A fixed value of a parameter is fixed which will be called a homogeneity factor 6. A pixel region is said to be homogeneous if, for all the pixels of this region, the following relation is satisfied: I (P (n)) I (P (n + 1)) <6 (1), with I (P (n)) and I (P (n + 1)) representing the light intensities of pixels P (n) and P (n + 1) adjacent neighbors, of rank n and n + 1, respectively, and n being an arbitrary integer. When the digital image has more than one color channel, the relation (1) must be checked for

all channels.

  These precisions being given, we will now describe the location phase of the objects present in a digital image. Naturally, it is assumed that, in a preliminary phase, a digital image of any resolution has been acquired. This image will be referenced below Img. Conveniently the digital image Img is stored in a storage device, for example the RAM memory of an information processing system, during the whole process of locating the objects (in the first phase) and extraction of these objects (during the second phase). This digital image Img serves as input to a system for locating and / or extracting objects in

  a digital image which will be described below.

  The first step of the first phase consists, in application of the diffusion, in segmenting the digital image Img into homogeneous regions, corresponding to the definition that has just been given. Figure 1A illustrates the digital image Img, segmented into homogeneous regions. During this step, i thus define i distinct regions, i being such that: 1 <i <mmax (2), with mmax equal to the maximum number of homogeneous regions found. The maximum number of homogeneous regions mmax, for a given digital image Img, naturally depends on the choice made for parameter 6, which constitutes a threshold value. For each region, we keep the intensities

  luminous maximum and minimum pixels of this region.

  The color range of the region is defined as follows:

interval [minimum; maximum] (3).

  At the end of the first step, the data acquired is recorded in a memory, advantageously in the form of lists, and in particular the data characterizing each pixel: identification of the region of belonging, coordinates, etc. In the second step, each of the regions is associated with an attribute that will be referred to hereinafter as "tag". More precisely, this attribute or label can take three values, which will be arbitrarily called "Not Labeled" or "NE" status, "Not Significant" or "NS" status, and "Significant" or "S" status. The state "NE" is an initial temporary state and the other two

  states, "NS" and "S", correspond to the regions of the rear

  plane and the regions of the objects to be located, respectively.

  By default, all regions are initialized to the "NE" state, as illustrated in FIG. 1A: referenced regions R (NE). This means that the regions have not yet been assigned a "real" state, even a temporary one. The coordinates of the pixels of each region, their index i (see (2)), as well as their label ("NE" in this case)

are stored in a memory.

  The process of locating objects in the digital image Img is started by arbitrarily initializing all image edge regions in the "NS" state. Figure 1B illustrates the labeling of the image at this stage of the process. There are now two types of regions, those at the edge of the image, which have been referenced Rm (NS), and those in the center of the image Img, still unlabeled, which have been referenced Rn ( NE), where n and m are integers such that: n + m = nmax (4) In Fig. 1B, the Rn (NE) regions have been left blank and the Rm (NS) regions are shown as hatched. Then, the process will continue, with a view

  to label all regions of the digital image Img.

  To do this, the entire digital image Img is scanned, starting, advantageously, from one of the ends and following a spiral path T composed of rectilinear straight line segments parallel to the edges of this image, as illustrated by FIG. . All the regions present in the digital image Img must be scanned, ie Nmax regions. To do this, it is sufficient that the pitch of the spiral T is sufficiently low. At the limit, the pitch may be equal to the distance separating two pixels, which can be determined easily knowing the resolution of the image

Digital Img.

  When the scan passes through a region, there are a priori two possibilities and only two, since the border regions of the digital image Img have been initialized to the "NS" state: a / the region is in the "Significant" state "S", or

  b / the region is in the "Not Significant" "NS" state.

  S15 In the first case, neighboring regions are tested for labeling. For each neighboring region in the state "NE", if the interception of the color ranges (see relation (3)) of the region "S" and of the neighboring region exists, then this neighboring region is also labeled state "S". Otherwise, we label

  this neighboring region in the state "NS".

  In the second case, neighboring regions are also tested for labeling. For each neighboring region in the "NE" state, if the interception of the color ranges of the "NS" region and the neighboring region exists, then this neighboring region is also labeled in the "NS" state. If not, label this

neighboring region in state "S".

  Figure lE illustrates an intermediate stage of the process of labeling regions. There are now three types of regions, those labeled in the "Not Significant" state, referenced as Ry (NS), those labeled in the "Significant" state, referenced as Rz (S), and those in the center of the image. Img, still untagged, referenced RX (NE), with x, y and z integers such as:

x + y + Z = nmax (5) -

  In FIG. 1E, the regions Rz (S) are filled with points to differentiate them from the two other types of

  regions (white and hatched patterns, respectively).

  Figure 1F illustrates the end of the labeling process: all regions are labeled definitively. Again, there are only two types of regions: regions in state "NS", referenced Rp (NS), and regions in state "S", referenced Rq (S), with p and q of integers checking the relation:

p + q = nmax (6).

  The data acquired during this step are also stored in a memory and complete those

  recorded at the end of the first stage.

  The third step of the first phase is to merge all neighboring regions into the "S" state. Indeed, as has been indicated, the regions in the state "S" delimit areas of the digital image Img corresponding to the objects to be located. In other words, each merge of neighboring regions determines a particular area of the image

  corresponding to one of the searched objects.

  In the illustrated example, as shown more particularly in FIG. 1G, there are four distinct significant areas, corresponding to as many localized objects. These zones, referenced Z1 to Z4, respectively, are marked with crosses in FIG. They are separated by other zones, constituted by the p Rp (NS) regions, that is to say in the "NS" state. All of these latter areas form the background of the image

Digital Img.

  The data recorded during the second stage

  are refreshed, to account for the aforementioned fusion.

  The fourth step of the first phase consists in determining, from the stored data, windows or boxes including, each, one of the localized objects, namely four boxes in the example described, EB1 to B4, respectively. The boxes, B1-B4, are advantageously of rectangular shape and voluntarily extended, to facilitate the steps of the second phase, or extraction phase of the actual objects. Obtaining the boxes is simple, since the coordinates of the pixels included in the merged zones Z1 to Z4 have been recorded. Just look for the coordinates of the highest pixel and left of these areas and the lowest pixel and right, for example. The determination of these two points is sufficient to fully characterize a rectangle. The rectangle obtained is then extended by adding a predetermined distance (expressed in pixels), along the vertical and horizontal directions. The data identifying the boxes

are saved.

  The second phase is performed for each of the significant areas Z1 to Z4, and comprises three steps. The input parameters consist of the digital image Img and each of the boxes, EB1 to B4, and their contents. Each of these boxes, B1 to B4, can be likened to an image

  elementary containing a single object to extract.

  The first step is to acquire a game of

Interception vectors.

  An intercept radius is defined as being a linear path along a line or column of a digital image, for detecting the abscissa or ordinate of an edge point of the image by relation to orthonormal reference axes. In the context of the method of the invention, the orthonormal axes are constituted by the orthogonal sides of one of the aforementioned boxes, B1 to B4, and the image is constituted by the significant zone that it encompasses, Z1 to Z4, respectively . We can therefore know the coordinates of each of the points of the contour of these zones, either with respect to the sides of these boxes, B1 to B4, or with respect to the sides of the digital image Img, by simple transformations of coordinates, the positions relative boxes, B1 to B4, within the frame of the digital image

Img, being known.

  More specifically, an interception ray run, said vertical, corresponds to launching, for each column of one of the boxes, B1 to B4, an interception radius (in the up-down direction or the down-axis direction). high) to detect the ordinate of a contour point of the corresponding enclosed area, Z1 to Z4. The ordinates of the detected points are stored in an interception vector, the size of which is equal to the height of

  the source image, that is to say one of the zones Z1 to Z4.

  Similarly, an interception ray ray, called horizontal, corresponds to launching, for each line of one of the boxes, B1 to B4, an interception radius (in the left-right direction or the right-left direction) to detect the abscissa of a contour point of the corresponding enclosed area, Z1 to Z4. The abscissae of the detected points are stored in an interception vector, whose

  size is equal to the width of the source image, that is,

say one of the zones Zl to Z4.

  In the context of the process of the invention, four ray throws are made: two vertical throws and two horizontal throws. Furthermore, as illustrated more particularly in FIG. 3A (taking as an example a box Bj, of any index j, and the enclosed area Zj), preferential procedure is carried out in the following order: Start 1: throw horizontal direction left- right; Throw 2: throw vertical up-down direction; Start 3: throw horizontal direction right-left;

  Throw 4: throw vertical up-down direction.

  The arrows shown in bold in Figure 3A

  represent the directions of the throws.

  One of the advantages of the method according to the invention is that it is not necessary that the background of the image, in this case the elementary image constituted by the inside of the box Bj, be uniform (transparent or a predetermined homogeneous color). On the contrary, the substance is not generally uniform but consists of a

  or more regions labeled "NS": Rp (NS) regions.

  Still in the general case, the intensity of the pixels encountered during a throw does not remain constant in the "NS" regions to abruptly pass to another value when crossing the boundary (contour) of these regions and

zone Zj (region "S").

  According to a preferred embodiment of the invention, an adaptive method is used to detect the intercept of the Zj contour. This method is going to be

  detailed with reference to FIG. 3B.

  We choose a predetermined integer number of pixels N. typically 3 to 5. We will consider, to illustrate the method, a horizontal throw axis 0X (throw 1), parallel to the horizontal sides of the box Bj (left-right direction) . We call C the current point of the progression of the throw along this axis 0X. Initially, C is the abscissa of the first pixel studied, that is to say the first pixel P (C) from the left vertical edge of the box Bj If we finally call S a set of N successive pixels to C when traversing the radius, C is a contour point if and only if, for any pixel P (x) belonging to S, the following relation is satisfied: I (P (C)) - I (P (x)) > 6 (7), with 6 representing the previously defined homogeneity factor (see relation (1)), and I (P (C)) and I (P (X)) being the intensities of the corresponding pixels. If the relation (7) is satisfied, all the pixels of S are heterogeneous with C, and C is, therefore, a significant contour pixel,

that is, an outline of Zj.

  In the opposite case, the next point with respect to the considered set S becomes the new point C, and the

  homogeneity test just described is repeated.

  It must be clear that for a given throw along an axis, the test may be negative over the entire width (or length) of the box B. This is the case, for example, of horizontal throws, in the upper part of Figure 3A, because the radius does not intercept the zone Zj. In this particular case, the last pixel encountered constitutes the point

of outline.

  In FIG. 3A, the path of the radius along the axis OX has been subdivided into several sections that illustrate the main cases encountered, except for the particular case described above (no interception). First, over a length 11 (expressed in pixels) greater than S (N pixels), ray tracing progresses in regions labeled "NS". The relation (7) is never verified. Then, it intercepts in XB a point which is

  at a point of contour, because relation (7) seems to be verified.

  However, the path length (in pixels) during which this relation could be verified is equal to 12, with 12 <S. This is actually a local noise or "artifact" that must be "jumped" by the radius firing, on pain of an erroneous coordinate detection of Zj (in this case an erroneous abscissa point). After passing through this noisy area, the ray tracing progresses again into regions labeled "NS", over a length 14. Finally, in Xj, ray tracing actually intercepts the contour of Zj. The relation (7) is satisfied because the length 14 (in pixels) of the zone encountered is greater than S. Xj is therefore the abscissa of one of the points belonging to the desired contour of the zone

significant Zj.

  The second step of the second phase consists of constructing an outline of the object to be extracted, which will be called primary, or raw, and which will be referenced Cop. This one is built from the interception vector set that has just been acquired. As mentioned previously, two ray throws are taken into account

  horizontal interception and two vertical throws.

  More precisely, the primary contour Cop is constructed by first analyzing the vectors in pairs: a pair of horizontal vectors, then a pair of vertical vectors, for example. This analysis is then used in a detailed way in a

  global study of the four vectors.

  A simplified method to ensure that a given point I belongs to the desired contour, that is to say is

  said "eligible", will now be detailed.

  We call EP (k) the set of points belonging to the primary contour of the object to be extracted (that is to say the significant zone Zj, for example), with k being an index associated with each point. The index k is an integer and the total number of points in the set is finite, because the image Img and each object to be extracted are of a numerical nature. The set EP (k) is initially empty. We call G (k) and D (k) the intercept vectors corresponding to the horizontal throws of rays, of left-right direction and

right-left, respectively.

  For everything checking the following relation: 0 <y <height of the image (8), the point I, of abscissa G (y) and of ordinate y, is eligible to belong to the primary contour if and only if: G (k) <D (k) (9) Similarly, the point I, of abscissa D (y) and y-ordinate, is eligible to belong to the primary contour, if and only if:

D (k)> G (k) (10).

  The coordinates of the points of the set EP (k), as well as the associated indices, are stored. This set can be represented by a polygon, and displayed on a

computer screen for example.

  However, the primary Cop outline has a number of imperfections or artifacts that must be removed. This state of affairs will be illustrated by reference to FIGS. 4A to 4D. These figures relate to a digital image Img having only one significant object referenced Oh. More precisely, the digital image Img is constituted by a photograph of the "identity" type representing the top of a subject: shoulders and face. It is assumed that this subject wears a dark Ve jacket, that his hair Ch is dark, and that the background (bottom) Ap is light in color. Finally, it is assumed that the texture of the background Ap is close to that of the face Vi of the subject, the face Vi being also of light shade. There is

  so a priori little bright contrast between the rear

Ap plan and face Vi.

  Under these conditions, it can be seen that the primary contour Cop (shown in bold lines) does not exactly follow the real contour of the object Ob, as shown more particularly in FIG. 4A. It comprises parasitic lines Rp, on the one hand, and a zone of erroneous contour Zce, on the other hand. The parasitic lines Rp are essentially due to very large residual noise noises, which could not be eliminated during the first step of this extraction phase (see Figure 3B). The erroneous Zce contour is due to the similarity of hue and

  texture between the face Vi and the background Ap.

  More precisely, in the example described, taking into account the particularities and the disposition of the different zones of the detected object Ob in the image Img, it is essentially the horizontal throws that will generate contour errors. Indeed, there is a significant contrast between the Ap background and the hair Ch, on the one hand, and between the background Ap and the jacket Ve, on the other hand. The analysis of the vertical shots will therefore not generate out-of-line errors since they intercept the boundaries of the hair Ch or the jacket Ve when they enter the region of the image Img covered by the object Oh. More specifically, in the example described, the Zce contour error zone is located in the lower face Vi, between the top of the jacket Ve and the bottom of the hair Ch. There is a multitude of horizontal slots forming back and forth on both sides of the face Vi, on the right and left of it, that is to say penetrating inside the area covered by this denier and emerging in the area covered by the Ap background. These slots

  may overlap with each other.

  FIG. 4B shows the primary contour obtained Cop, isolated from the object Ob, that is to say from the photographed character. It is easy to understand that such an outline is not

not exploitable directly.

  The third step of the second phase consists of the reconstruction, from the primary contour obtained in the previous step, of a Cor object contour that can be called real, with very great precision. This step

  is illustrated with reference to Figures 4C and 4D.

  To do this, we apply a set of rules, which

will translate into sub-steps.

  In particular, it is necessary to delete

  exist, the recoveries of the outline.

  Similarly, any contour discontinuities due to poor digital image quality should be detected and the polygon mentioned above should be smoothed.

  respecting the principle of continuity of curves.

  Eliminating overlaps from the polygonal contour

  will now be explained with reference to Figures 4E to 4G.

  FIG. 4E shows two contour sections, which are in the form of two sequences of line segments. These two sequences are associated with the indices a and b respectively, and we will arbitrarily call the corresponding sections right and left contours. There are five inflection points for each of the two contours, Pa (1) to Pa (6) and Pb (1) to Pb (6), respectively. It can be seen that the two contours intersect at two points (two intersections), I and J, because the point

  Inflection Pb (4) is located to the left of the left contour.

  The two intersection points, in the example, surround the

point Pa (4) of the left contour.

  It is therefore necessary to eliminate the recovery. To do this, for each intersection detected, I and J, we proceed

as described below.

  By way of example, we consider the segments [Pa (3), Pa (4)] and [Pb (3), Pb (4)] associated with the point of intersection I. For each of these segments, measures a parameter, Ca and Cb respectively, reflecting a continuity of the segments with respect to the contours to which they belong. For the segment [Pa (3), Pa (4)], the parameter Ca obeys the following relation (11) Ca = Pa (2) Pa (3) FP (4) mod n) + (Pa (3) Pa (4) Pa (5) mod 7r), with "Pa (x) Pa (y) Pa (z)", an oriented angle formed by any three consecutive points, x, y and z of the polygon, and "mod n" , an operation modulo z. For the segment [Pb (3), Pb (4)], the parameter Cb obeys the following relation (12): Cb = (Pb (2) Pb (3) Pb (4) mod n) + (Pb (3) ) Pb (4) Pb (5) mod 7t) The Ca and Cb parameters reflect what can be called a "degree of continuity" with respect to the polygon, the lower these values are, the greater the degree of continuity

is strong.

  It follows that, in order to eliminate the detected overlap, the least "continuous" segment, in this case the segment [Pb (3), Pb (4)], is modified because Cb is more important than Ca (in FIG. 'example). Changing this segment means that you must delete one of the two endpoints in this segment. To determine the point to be removed, Pb (3) or Pb (4), proceed as illustrated in FIG. 4F. The orthogonal projections H (3) and H (4) of the points Pb (3) and Pb (4), respectively, ie the distances H (3) Pb (3) and H (4), are calculated. Pb (4). In the example described, H (4) Pb (4) <H (3) Pb (3) is obtained. It is concluded that it is point Pb (4) that must be deleted because the most

close to the segment [Pa (3), Pa (4)].

  The two points Pb (3) and Pb (5) are then directly connected as shown in FIG. 4G (segment in bold lines). Crosses represent the deleted segments: [Pb (3), Pb (4)] and [Pb (4), Pb (5)] of the right contour. It can be seen that there are no more intersections between the two partial contours, the intersection J having

  also been removed by this process.

  We will now describe the smoothing process of the polygon. To do this, we pass in parameter three integers, expressed in pixels, namely a length L, a height H and a distance parameter E. To fix the ideas, we can select the following typical values: L = 5 pixels, H = 5 pixels and 5 <E <10.pixels. The input parameters, H and L, define a rectangular virtual box Bv. Then, iteratively (stepwise) is performed a series of operations consisting in following the contour of the polygon and selecting, at each step, three consecutive points of the polygon (that is to say two equally consecutive segments ). FIG. 4H shows a polygon section comprising six points of inflection, P1 to P6. For each series of three consecutive points, for example the series of points P2 to P3, the

  dimensions in pixels of a box B enclosing them, that is to say

  say in this case a rectangle of height h and width 1. If 1> L and / or h> H, we mark the intermediate point, P3, to a first state said invalid for the polygon. Its coordinates are stored, but this point is not immediately deleted. In the opposite case, the intermediate point, P3, is marked at a second state said valid. At the end of the iteration process, all the points

  invalids of the polygon are thus known and memorized.

  The process therefore amounts to comparing, at each new step, the width L and the height H of the real box B enclosing three consecutive points, at the width 7

  and the height h of the virtual box Bv.

  The third parameter, the distance E, will be used to effectively remove invalid points and smooth the polygon. FIG. 4I shows a polygon portion comprising fourteen points of inflection, P'F1 to P'14. With respect to the vertical (in the example described), there are three distinct parts: two extreme parts comprising the points P'l to PF4 and P'11 to P'14, substantially aligned, and an intermediate part comprising the points P '5 to P'9, strongly shifted to the left. The points P'5 and P'9 were represented in black on the

  Figure 4I and are points of the polygon declared invalid.

  Indeed, because of the proper length of the segment [PF4, P'5], the box B 'encompassing the points P'4, P'5 and P'6 has at least a width greater than the predetermined width L , entered as input parameter. The same is true of point P'10, for the same reasons. However, if we stick to the criteria previously used (width L and height I), the intermediate points, P'6 to P'9, will be declared valid. The distance e between two invalid points, in this case P'5 and P'9, makes it possible to determine the number of valid intermediate points, ie four in the example: P'6 to P'8. If e is greater than the predetermined distance, E, entered as a parameter, then the intermediate points P'6 to P'8 must be deleted. We also delete the invalid points, P'5 and P'9, and we gather the valid points upstream and downstream of the treated section, P'4 and P'11. A new segment is obtained, represented in bold dashed lines in FIG. 4I, directly connecting the points PF4 and P'11. The reading operation is now complete, at least for this part of the polygon. Figure 4C illustrates the actual Cor contour finally obtained and the Ob object it surrounds. Figure 4D illustrates the actual contour only Cor. We notice the disappearance of parasitic lines (Figure 4A: Rp) and the actual contour obtained Cor

  faithfully follows the limits of the Ob object.

  From the moment when the real contour Cor of the object Ob to be extracted is reconstructed, the object itself Ob can be obtained easily from the source image Img, excluding all the points situated outside the contour and keeping only the information data associated with the points inside the Cor contour: brightness, colors, coordinates, etc. This information can be obtained from the overall digital image Img previously stored. This digital image Img constitutes one of the input parameters, as has been indicated. These information data can be memorized in order to reconstitute at any time the extracted object Oh, for later uses.

which will be detailed below.

  By way of non-exhaustive example, the application of the method according to the invention will be briefly described in the so-called "compositing" technique. This is to superimpose, for special effects for example, the object Ob, or foreground, on a background entirely different from the original background: landscape, crowd, etc. The first step is to obtain, following the phases and steps of the method of the invention, the final contour of the object Oh, which has been called real Cor (Figure 4D). This Cor contour can then be filled with a uniform color, for example

  example the black, to constitute a mask of transparency.

  The superposition of the original digital image Img, the transparency mask obtained in the previous step and an image constituting the new background will make it possible to obtain a new image on which the foreground is the object. Background

and the background the new background.

  For some applications, it may be desirable for the Ob object to show more or less pronounced the new background. This result can be obtained by well-known techniques, by assigning to the mask a determined opacity coefficient. Similarly, it is possible to obtain several layers of

  superposition, or "layers" according to the English terminology

  Saxon, by repeating all or some of the previous steps.

  In general, all types of special effects can be applied to an image represented by the cut-out object alone or on a composite image, of the type of

  which has just been described: video inversion, pseudo

  solarisation, fuzziness, various deformations, etc. These methods are well known to those skilled in the art, and they are

no need to describe them further.

  Other conventional types of exploitation of digital objects, such as those extracted by the method of

  the invention, relate to indexation, post-

  processing, archiving, classification, etc. If the digital image Img has several distinct objects, as has been implicitly accepted so far (FIG. 1H: zones Z1 to Z4), steps 1 to 3 can be repeated if it is desired to extract all the objects. We then have a list of objects present in the image, each object being associated with a polygonal set of points which delimits the contour in the original digital image Img. Otherwise, the object extraction process itself, in accordance with the second

phase of the process, is finished.

  As an alternative, one can provide, at the end of the first phase, a manual step of selecting, for example after display on a computer screen, one of the significant areas, Z1 to Z4, or more simply one boxes B1 to B4, using a pointing device, the "mouse" type or other, by clicking on one of these boxes. According to this variant, only the object included in the selected box is extracted, leaving the three stages of the second phase

fully automatic.

  According to the preferred embodiment of the method according to the invention which has just been described in detail, two successive phases are carried out: location of object contour and extraction. However, as was also stated in the preamble of the

  this description, according to another embodiment, the

  locating phase can be omitted.

  First, if we know, a priori, that the digital image has only one object to extract, the first phase can be omitted without particular difficulties. In this case, it is the frame of the Img digital image itself that can serve as a selection box. The coordinates of the desired outline are

  determined from the sides of the digital image! mg.

  If the digital image Img has several different objects, the steps of the first phase being omitted, it is not possible to discriminate them automatically. It is then necessary to define "by hand" a selection box surrounding each object to be extracted. To do this, the Img image can be displayed on a screen. An operator visualizes the objects and defines one or more boxes including, each, one of the objects he wishes to extract. Indeed, it is not necessary for all objects to be surrounded by a selection box if the aforementioned operator is only interested in part of the displayed objects. The

operating method is simple.

  A selection box can be defined by an interface called "ManMachine" ("HMI") and displayed on a computer screen. The operator has a keyboard and a pointing member, for example of the mouse type, advantageously comprising two function buttons, which will be called right and left. The operator requests the interface to display, superimposed on the image to be processed Img, a rectangle polygon of predetermined standard size on a computer screen. This request can be made using a menu, a keyboard shortcut or by clicking on an icon of a toolbox also

displayed on the screen.

  By clicking inside the standard selection rectangle, using the left mouse button, and holding down this button, the operator can move the

  selection rectangle polygon on the screen surface.

  The rectangle is erased from the original position and is re-

  displayed in the last position reached.

  By clicking inside the selection rectangle with the right mouse button (usually assigned to so-called contextual functions), and holding down this button, the operator can change the length and height of the rectangle polygon. Releasing the right button

  starts the steps of the object extraction phase. Those-

  They then execute automatically, as in the case of the preferred embodiment, and there is no need to redescribe them. In another variant, the two operations, modification and launch of the extraction phase, are combined. The polygon resulting from the extraction phase can be memorized and possibly displayed on the screen. It is also quite possible to define several selection polygons on the screen. The steps of the extraction phase are then executed for each polygon of

  selection sequentially.

  So far, we have admitted, at least implicitly,

  that the digital images Img to be processed were static.

  The method is, however, entirely compatible with the processing of dynamic digital images, that is to say images forming a video sequence. It is enough that the

  system for implementing the method which will be described below.

* after is fast enough to perform in particular the successive acquisitions and the location and / or extraction of the objects contained in each of the images presented to it, as well as the intermediate and final data storage. In addition to the attributes associated with the different objects of the same image Img, it is possible to add to them an additional attribute making it possible to identify the image or images of the aforementioned video sequence in which

they are present.

  In a further variant, it is possible to carry out an additional phase comprising a step of comparing the objects successively acquired, followed by a step of eliminating the objects repeating themselves identically in different images of the video sequence, adjacent or not . Naturally, it is possible to take into account possible image modifications by simple homothety or variations of various other parameters (brightness, etc.) lower than a predetermined threshold. This arrangement avoids redundancies, in particular makes it possible to record only once the bulk of the data relating to an extracted object. We will now describe a system for location and / or automatic extraction of objects in a digital image, implementing the method according to the invention,

with reference to Figures 5 and 6.

  Figure 5 schematically illustrates a complete chain 1 of processing and digital image exploitation. It comprises a device 10 for acquiring one or more analog images, the output of which feeds a

  analog-digital converter 11.

  For example, the source of analog images

  can be constituted by a conventional video camera.

  These first two devices can be omitted if one or more sources directly generating images in digital form are available: camera and digital cameras, scanner, etc. The outputs of the analog-to-digital converter 11 and / or the aforementioned apparatuses, sources of digital images, feed a digital image storage device 12. The latter presents these digital images at the input of the localization and / or extraction system objects in an image

digital 13 according to the invention.

  Finally, once extracted from the digital image presented at the input of the system 13, the digital object or objects are transmitted to one or more operating devices 14, for archiving, performing special effects or effects. special, etc., according to the

targeted applications.

  FIG. 6 illustrates in greater detail an exemplary embodiment of the system for locating and / or extracting objects in a digital image, a system that

it has been referenced 2.

  The system 2 comprises a device for location and / or automatic extraction of objects proper 20, implementing the method according to the invention. In the example described in FIG. 6, the member 20 is under the control of recorded program information processing circuits 21. In a conventional manner, these circuits 21, called "Central Unit", or "CPU" I, are advantageously made on the basis of a microprocessor, associated with various conventional circuits (not shown): read-only memory for the recording of the operating system and various routines and firmware, RAM for the temporary recording of data and programs,

  buffers (or "buffers" according to the English terminology

  Saxon), etc. The central unit 21 is also connected to a mass memory (hard disk) 23, for the longer-term recording of data and / or programs, to input and data input devices and / or to program instructions, such as a keyboard 24, pointing devices (for example a mouse 25) and various recording / reading units, internal and / or external,

  represented under the unique reference 26 (reader /

  floppy or magnetic tape recorder, CDROM reader, etc.), as well as output devices, such as a display screen 27 (via a non-graphic card

  shown), a printer 28 (via the input circuits-

  output 22), or a plotter (not shown).

  The system 2 finally comprises input-output circuits 22, placed under the control of the central unit 21, which can

  include in particular an input-output card (not shown), provided with various ports of serial, parallel, "SCSI" (for "Small Computer System Interface"), "USB" (for "Universal Serial Bus"), etc. ., a digital image acquisition card and / or an analog-digital conversion card, if necessary. This

  last card plays the role of circuits 11 of Figure 5.

  The input-output circuits 22 are intended to acquire, or at least receive, the signals representing the digital image or images containing digital objects to be located and / or extracted automatically. They receive as input one or more image sources, and in particular, in the example described in FIG. 6, the outputs of a digital camera 3, an analog or digital video camera 4, and a video camera. scanner 5. It is also expected that the digital images can be transmitted by a local (not shown) or remote network, via a telephone line, dial-up type or digital type ("ISDN" I), and a modem 29. It can This includes images from Internet sites transmitted over the Internet R1, as suggested by

figure 6.

  It should be understood that the digital images can also be input into the system 2 via media (such as floppy disks 260, Cédrom, magnetic tapes) on which they were previously recorded, in an appropriate format ("JPEG", " BMP "

"TIF", etc.).

  Whatever the source, once acquired, and possibly converted to a format compatible with the system 2, the digital image or images are presented to the central unit 21 to be stored temporarily in its RAM (not shown) . The corresponding digital signals can also be stored on the hard disk 23, for later use. The display member 27 may finally display the image being processed. Under the guidance of an operator (not shown), entering instructions via the keyboard 24 and / or the mouse 25, and under the control of recorded programs, the signals associated with a digital image to be processed are then transmitted to the object locating and / or extracting member 20, under the control of the central processing unit 21. The digital signals serving as input to the method according to the invention, as has been described with regard to FIGS. 1A to 4D, come directly from the central unit 21 (where they are stored in its random access memory) or

  transit through it from the hard drive 23.

  The member 20 for locating and / or extracting objects in a digital image can be in various configurations. It can be of a fully wired type to obtain maximum speed, that is to say it is in the form of a PLC receiving the image to be processed and outputting digital signals representative of the object or objects ( s) extracted (s) signals transmitted to the central unit 21, for example for storage in the hard disk drive 23, and / or directly to an operating device 14. The disadvantage of such a

  solution is a high rigidity of the system.

  In another variant, only repetitive routines are wired, or more precisely recorded in read-only memories ("ROM" or "Read Only Memory"), possibly reprogrammable ("PROM" or "Programmable Read Only Memory", whatever 'is the type:

"EPROM", "EEPROM", etc.).

  The various steps of the method according to the invention take place under the combined control of the instructions received from the central unit 21, that is to say those of a program loaded in RAM, instructions and parameters entered by the operator and micro-instructions of the aforementioned routines. The aforementioned program can be loaded initially or updated using a medium such as a floppy disk 260 (or a CDRom, etc.), read by the unit 26, or transmitted by telematic means ( not shown) or download from the network

Internet RI.

  It is also possible, in the location and object extraction circuits 20, a fast processor, specialized in the processing of digital signals ("signal processor" according to the English terminology), or co-processor, having advantageously an architecture called "RISC" ("Reduced Instruction Set Computer" or processor with reduced instruction set). This processor

  additional unloads the main processor, ie

  say that of the central unit 21, instructions more specific to the method of the invention. These are the instructions or no program needed to obtain the different phases and steps of the process: segmentation of the digital image, image labeling, etc. In all cases, the object locating and extracting device 20 comprises in particular intermediate storage circuits, such as registers, clock circuits and interface circuits, at least with the central unit 21 In the usual way, the exchanges are carried out via a system bus. The circuits 20 may also have a direct access ("DMA", according to the English terminology) to the RAM of the unit

Central 21.

  In yet another variant, the circuits of the member 20 can be placed directly on the graphics card or on a "daughter" card associated with this card, the graphics processor with which such a card is also being used for the specific needs of the card. process according to

the invention.

  The digital signals generated at each instant, during the various steps of the method, can be stored either temporarily locally, ie in the member 20, if it is provided with a sufficient random access memory, or transmitted to the central unit 21 for storage in its own RAM and / or on the hard disk 23. The various transformations of the digital image to be processed, as well as the progress of the location and extraction of the objects contained in it, can be viewed on the screen 27, in one or more windows. Although the process, in a preferred embodiment of the invention, may be fully automatic, an operator may be offered the opportunity to intervene at various stages of this process. It can enter, for example via the keyboard 24 or the mouse 25, instructions and initialization setting of the system 2, indicating in particular which is the source of images (scanner 5, etc.), which is the format of the images to be processed, what pretreatment must be done (eg analog-to-digital conversion or format conversion), which mode of implementation of the process is to be implemented (automatic location and extraction of objects in the preferred mode in accordance with the invention), specify how and where to record the digital signals representing the extracted objects, etc. During the process, the operator can also intervene to set certain parameters, in particular to manually delimit extraction zones (if only this phase must be performed), to select a part of the extracted objects automatically, or even only one of objects, what will then be the exploitation of the objects extracted, etc. To do this, instructions can also be transmitted to the operating device 14 via the central unit 21 and after manual input. The device 14 can also be provided with own commands (not shown), as well as a display screen 140 (television screen, etc.), and more generally, all the devices that are necessary for it to perform a post -treatment on the digital objects transmitted by the system 2, accompanied or not by

original digital images.

  Advantageously, to enter and enter the above instructions and / or parameters, the operator may have drop-down menus or the like presented to him on the screen 27, under the control of the central unit 21. to "click" on the desired options, prior to the progress of the phases and steps of the method according to the invention, during the process of locating and extracting objects, and / or at the end of this process, in particular to specify the destination and the nature of the post-processing to be performed on the extracted objects. If the system is not directly connected to the operating device 14, the data relating to the objects can be recorded for further processing, first locally, on the hard disk 23, but also on various media: floppy disk 260 (or magnetic tape, CedeRom, etc.). They can then be transmitted and used later, advantageously with the

  data corresponding to the original digital image Img.

  These same data can also be transmitted by any telematic means, even by the Internet network RI, and be stored in a remote place or be used in time

almost real in this place.

  The printing means 2, or the like (plotter, etc.) can also be used to obtain a

  graphical representation of the objects extracted,

  only to perform the data listing

associated.

  In another variant embodiment, the system for locating and / or extracting objects in a digital image Img can be integrated in the operating device 14 and operate completely autonomously, without any outside intervention, with the exception the initial loading of a program (which can also be resident in ROM) and a particular parameterization. As examples, the complete image processing chain of the figure ??? could be summarized as a digital video camera (Fig. 6: 4) and the operating device 14 could be provided with a clean display screen 140. The device 14 incorporates the system 2 and receives the video images produced by the camera. The device 14 is programmed to produce a fixed post-processing on the extracted objects and the final image is displayed on the screen 140. On reading the above, it is easy to see

  that the invention achieves the goals it has set for itself.

  The method and the system for implementing the method according to the invention make it possible, in a preferred embodiment, to fully automatically process a digital image in order to extract all the objects.

  digital images contained in the aforementioned digital image,

  that is to say without external intervention, whatever the nature of the background (background). The latter does not need to be uniform. The object detected and / or extracted is very precise (to the pixel, in the form of polygons), and quickly, even if the background above

is noisy.

  It should be clear, however, that the invention is not limited to only the examples of embodiment explicitly described, especially in relation to FIGS. 1A to 6. In particular, the boxes encompassing the significant regions could differ from the rectangular shape, although the latter is particularly adapted to serve as reference axes and to determine the coordinates of the

covered points.

  It must also be clear that, although particularly suited to applications of the type of special effects, the invention can not be confined to this type of application. It applies whenever you want to automatically extract an object from a digital image.

RE VE N DICATIO N S

  A method for processing a digital image (Img) for extracting digital objects constituted by homogeneous optical characteristic pixel regions, said regions having a closed contour, characterized in that it comprises: phase of locating at least one of said digital objects (Z1-Z4), comprising at least one step of creating in said digital image a closed area (B1-B4) encompassing this digital object (Z1-Z4), the coordinates pixels of the component being determined with respect to a determined reference frame related to said closed area (B1-B4); and an automated extraction phase of at least one digital object (Z1-Z4) comprising the following steps: a / acquisition of a set of interception vectors with said contour of the object (Z1-Z4) ) each representing a coordinate of a point of this contour with respect to said frame, the interception being determined by the prior fixing of a threshold parameter characterizing said homogeneity and the detection of the variation of this parameter beyond this threshold; b / the acquisition, from this set of interception vectors, coordinates, with respect to said reference, points describing a first contour (Cop), the digital object (Zl-Z4), said primary; c / reconstruction, from this so-called primary contour (COp), of a second contour representing the real contour (Cor) of said object to be extracted, by elimination of noise

  parasites and smoothing said primary contour (Cop).

  2. Method according to claim 1, characterized in that said digital image (Img) comprises at least one color channel, said homogeneous optical characteristics of the pixels are determined by testing, for each of said channels, whether the following relationship is verified for all the pixels: | I (P (n)) - a (P (n + 1)) <6, where 5 is said threshold and I (P (n)) and I (P (n + 1)) are the light intensities adjacent adjacent pixels P (n) and P (n + 1) of ranks n and n + 1 respectively, n being a

arbitrary integer.

  3. Method according to claim 2, characterized in that said object-locating phase (Z1-Z4) is also automated and comprises the following steps: the segmentation of said digital image (Img) into regions (Ri (NE)) ) consisting of adjacent pixels with homogeneous optical characteristics, said homogeneity being determined by the prior fixing of a so-called homogeneity parameter and the detection of the pixel-to-pixel variation of this parameter less than a second predetermined threshold; - determining for each of said channels a color range by retaining the maximum value and the minimum value of said light intensities, said range being defined by the relation: interval [minimum; maximum]; labeling of said homogeneous regions by means of an attribute representing the following states: initial state said "Not Labeled", said "Not Significant" state for the regions (Rx (NS)) not belonging to said object, and said "Significant" state for the regions (Rz (S)) belonging to said object, the labeling being effected by the initialization of all the regions situated at the edge of said digital image (Img) to said "Not Significant" state and, by scanning all said homogeneous regions, gradually assigning one of said "Significant" or "Not Significant" states, by testing the interception of said color ranges of two adjacent regions and assigning it to the region next adjacent state of the previous region if the test is positive and the opposite state if the test is negative; - merging neighboring regions, associated with said "Significant" status attribute, to form regions

  corresponding to said digital objects to be extracted (ZI-

  Z4); and said step of creating closed areas (B1-B4)

  encompassing digital objects (Z1-Z4).

  4. Method according to claim 3, characterized in that said scanning is performed following a spiral path (T) parallel to the edges of said image (Img), starting with said homogeneous regions (Rm (NS)), initialized to said state "Not Significant" and located at the edges, and ending at the center of said image (Img).

  5. Process according to claim 3 or 4,

  characterized in that said closed areas (B1-B4) encompassing said digital objects (Zj) are delimited each by a rectangle, the vertical and horizontal sides of which constitute said reference frame, in that said step of acquiring a game of interception vectors consists, for each of said areas (Bj), of analyzing the successive pixels composing them, from the four sides of said rectangle and on courses (throwing 1 - throwing 4) parallel to these sides, until interception with said contour of the encompassed digital object (Zj), so as to determine a set of four vectors representing the coordinates of the points of said contour with respect to each of the sides of said rectangle, and in that said interceptions are effective when the following relation is satisfied, for a predetermined number N of consecutive pixels on each of said four courses (throw 1 - throw 4): I (P (n)) - I (P (n + 1))> b, with 8 being said second d predetermined threshold, I (P (n)) and I (P (n + 1)) being the luminous intensities of adjacent pixels P (n) and P (n + 1), of rank n and n + 1 respectively, and n and N being arbitrary integers. 6. Method according to claim 5, characterized in that the determination of said closed areas (B1-B4) delimited, each by a rectangle, comprises the following operations, performed using a human-machine interface device comprising at least one screen (27), a keyboard (24) and a pointer (25), of the mouse type, provided with at least two function buttons: - the display on said screen, superimposed on said image numerically (Img) of a rectangle of standard dimensions, by actuation of said pointing member (25) or by use of the keyboard (24); selecting said rectangle by clicking on it by actuating one of said function buttons; moving said rectangle on the surface of said screen (27) by holding said first depressed button and moving said pointing member (25); modifying said standard dimensions by clicking inside said displaced rectangle, with the aid of a second function button, so as to encompass said digital object (Zj); and and launching said automatic extraction phase,

  by releasing said second function button.

  7. Method according to claim 5, characterized in that the step of constructing said primary contour (Cop) consists in comparing said vectors two by two, according to each of said vertical and horizontal directions, so as to obtain the abscissa and ordinate of each points of said primary contour (COp) relative to said vertical and horizontal sides of said rectangle, and to record the coordinates of all

the points thus determined.

  8. Method according to claim 7, characterized in that said step of reconstructing the real contour (Cor) from the registered coordinates of the points of said primary contour (Cop) consists in performing the following correction operations: - the suppression of the recoveries of said primary contour (Cop) along said vertical or horizontal directions; and smoothing the curve represented by all the points of said primary contour (Cop) by detecting discontinuities of said curve and eliminating them; and in that the new coordinates obtained for the points of said corrected contour constituting said contour

real (Cor), are recorded.

  9. Method according to claim 8, characterized in that said deletion of recoveries of said primary contour (Cop) comprises the following operations: detection of intersection points (1, J) between pairs of line segments forming said primary contour (COp); determining, for each of said line segments, a degree of continuity with respect to the line segments surrounding it, determining the line segment having the lowest degree of continuity; calculating the projections (H (3), H (4)) of the end points of this line segment on the other segment of said pair, determination of the weakest projected and elimination of the corresponding endpoint; and - and direct connection of the remaining end point (Pb (4)) to the next point (Pb (6)) of the contour 10. The method of claim 9, characterized in that said curve smoothing comprises the following operations: the selection of three determined input parameters, said height H, length L and distance E, expressing distances in pixels; tracking step by step of said primary contour, by selecting, at each step, three consecutive points (P2, P3, P4), ends of two equally consecutive segments; the calculation of the length 1 and the height h of a rectangular box including said three points (P2, P3, P4) and comparison with the lengths L and height H determined; the marking of the intermediate point (P3) of said three consecutive points according to one of the following two states and the recording of this state: a / invalid, if l> L and / or h> H, b / valid, in the opposite case; calculating the distance e in pixels from two consecutive invalid points (P'4, P'5) and eliminating these points and valid intermediate points if the following relation is verified: e> ,; and - direct interconnection of upstream ('4) and downstream points

(PI11) of said primary contour.

  11. System for implementing the method according to

  any one of the preceding claims,

  characterized in that it comprises input circuits

  an output (22) connected to at least one digital image source (3-5, 29), stored program data control and processing means (21), receiving signals representing said digital image acquired by said digital image circuits; input-output device (22), storage means (23) for reading and permanently recording digital data, in particular data corresponding to said extracted objects, and data acquisition means (24, 25) connected to said means for controlling and processing digital data (21), display means (27) and specialized data processing means (20) cooperating with said digital data processing and control means (21), performing said phases location and / or

extraction of objects.

  12. System according to claim 11, characterized in that said digital image source comprises a digital camera (3) and / or a camera

digital (4).