EP1053533A1

EP1053533A1 - Method for spatial segmentation of an image into visual objects and application

Info

Publication number: EP1053533A1
Application number: EP99901651A
Authority: EP
Inventors: Pascal Faudemay; Gwena[L Durand
Original assignee: Universite Pierre et Marie Curie Paris 6
Current assignee: Universite Pierre et Marie Curie Paris 6
Priority date: 1998-02-04
Filing date: 1999-01-28
Publication date: 2000-11-22
Also published as: FR2774493B1; WO1999040539A1; FR2774493A1

Abstract

The invention concerns a method for spatial segmentation of images into visual objects to obtain objects having semantic significance comprising steps which consists in dividing the image into initial regions, fusing the neighbouring initial regions according to certain similarity functions and obtaining enlarged images, then repeating said phase with the remaining initial regions and the enlarged regions until there is no possibility of further fusion. The invention also concerns the use of said method for characterising, classifying, storing, detecting objects, scenes, planes, groups of planes, movements, in multimedia documents, particularly of video.

Description

METHOD FOR SPATIAL SEGMENTATION OF AN IMAGE INTO OBJECTS

VISUALS AND APPLICATION

The invention relates to the field of analysis of the information contained in color images, in particular from multimedia documents, and in particular videos. This analysis is intended more particularly, but not exclusively, to allow indexing of audiovisual documentation.

The problem consists in particular in making a temporal division of video sequences into scenes which constitute narrative units of these sequences, with a view to the storage and the selective visualization of certain scenes by the users. Such a level of understanding is not directly accessible by known segmentation methods. Methods have been developed for segmenting grayscale images or for segmenting moving objects in the images of a video. But none of the color image segmentation algorithms provides sufficient results when the images to be analyzed are taken from "real life". The main methods can be grouped into pixel type, edge type, and region type methods.

In pixel segmentation methods, a region is defined as a connected component of a set of pixels specified by a membership function, possibly fuzzy, in a color class of the Color Space (abbreviated as EdC) . These methods are mainly differentiated by the way they define the color classes and the membership functions of these classes. The simplest of these methods is a rigid quantification of the EdC, as described in the articles by C. Carson, S. Belongie, et al. "Region Based Image Querying", Proc. CVPR'97, Workshop on Content-Based Access of Image and Video Libraπes, 1997, and by JR Smith and S -F Chang ^' Tools and Techniques for Color Image Retπeval ", Proc SPIE'96, Storage Se Retπeval for Image and Video Databases IV, Vol 2670, 1996 However, some similar colors in the 'EdC can be in separate classes

Another pixel type method is a histogram thresholding method, in which the peaks and valleys appearing in one or more histograms corresponding to the different dimensions of the EdC are sought. The detected valleys are then used as limits between the classes of colors (as described for example in the article by R Hayasaka, J Zhao and Y Matsushita, ^" Outstanding Object-Oriented Color Image Segmentation Using Fuzzy Logic" Proc SPIE'97 Multimedia Storage and Archiving Systems il, Vol 3229, 303-314, 1997

The clustering methods (clustering in English terminology) of the EdC are multidimensional extensions of the previous thresholding techniques and apply classification algorithms such as the algorithms for finding nearest neighbors, (see the article by R Ferπ and E Vidal, ^" Color image Segmentation and labeling through multiediting and condensing", Pattern Récognition Letters, vol 13, No 8, pp 561-568 1992), the algorithm of mean K (K-mean in English terminology) or fuzzy mean C (Fuzzy c-mean in English terminology) (see the article by YW Lim, SU Lee, ^" On The Color Image Segmentation Algoπthm Based on the Thresholding and the Fuzzy c-Means Techniques", Pattern Récognition, Volume 23, Number 9, pp 935-952, 1990) These algorithms make it possible to search for potential clusters of colors in the images Finally, certain methods seek the EdC allowing an optimal representation of the images using techniques such as analysis into main components or the Karhunen-Loève transform, such as that described in the article by SE Umbaugh et al, ^" Automatic Color Segmentation Algoπthm with Application to Skin Tumor Feature Identification ", IEEE Engineering in Medicine and Biology, Vol. 12, No. 3, pp. 75-82, 1993.).

The first drawback of these methods is that a robust search for peaks in the histograms or clusters in the EdC is not easy, in particular in the case of low-contrast images, and can be costly in computation time. In addition, these methods implicitly consider that if two pixels, that is to say two homogeneous image points, whether the support is a video image or not, belong to the same region of the image, their respective colors belong to the same color class or are close in EdC. This only applies correctly in clip-art or cartoon images, but generally not in real complex images. When extracting regions, these methods must therefore analyze the immediate vicinity of the pixels to determine to which region each pixel should be attached.

In addition, in the contour type methods, the contours are detected and used to determine the limits of the regions. However, edge extraction methods are not segmentation techniques by themselves and must be combined with at least one of the other methods. In addition, the contours obtained in the case of poorly contrasted or highly textured images are difficult to use since they are generally not closed.

In the region-type algorithms, a region is defined as a set of connected pixels satisfying a given homogeneity criterion, for example an area which contains only one color present in 95% of the pixels of the region. In the so-called split-and-merge techniques (split-and-merge in English terminology), the non-homogeneous regions are split into sub-regions, then analyzed recursively until the criterion of homogeneity be satisfied for each of the regions obtained. The neighboring regions thus obtained are grouped together.

The so-called region-growing techniques are region-type methods in which a certain number of initial zones, used as growth seeds, are first sought. The neighboring pixels and regions are iteratively included in these initial zones until a stopping criterion is satisfied, for example when the number of regions obtained is less than a threshold. A known example of this category of method is the so-called "topographic basins" algorithm in which an image is considered as a topographic relief, where the altitude of each point can, for example, be proportional to its light intensity. The bottom of the most important basins is pierced, and the relief is immersed in water. The pools gradually fill up, delimiting the main regions. This method is very sensitive to noise and costly in computation time. Another method in this category is the method based on recursive spanning trees of minimum weight (Recursive Shortest Spanning Trees, in English, RSST for short). RSST (see the article by O.J. Morris et al, in “graph theory for image analysis: an approach based on RSST”, IEEE proceedings, vol. 1333, 146-152, 1986) considers each pixel as an initial region. The regions with the closest average colors are merged recursively, favoring the merging of small regions. Most of the above methods operate at the pixel scale. This makes them particularly sensitive to local variations in intensity and therefore to textures.

This sensitivity is necessary for artificial vision or pattern recognition applications for which the extraction of exact contours is essential, but it is penalizing when looking for large semantically significant regions.

The invention aims to overcome these drawbacks by proposing a segmentation into regions of sizes such that the regions thus segmented still have a semantic meaning in the context of the image. By semantic or semantically significant object, it is understood an object corresponding to the real world, for example a face, a sky, etc. Several semantic objects can compose another semantic object (for example a hair, a face and a jacket constitute a person) hereinafter called composite semantic object. The semantic objects composing a composite semantic object can also themselves be composite semantic objects (for example the face is composed, inter alia, of a nose, a mouth and eyes).

The use of segmentation of images into objects with significant semantic value is a key step in the process of analyzing and understanding the content of multimedia documents, in particular video documents.

The invention makes it possible to segment the images into significant objects while neglecting the details. The invention thus aims to obtain a robust segmentation in the presence of possibly very textured images, and insensitive to insignificant details which could lead to unnecessary over-segmentation of large homogeneous regions, for example a black cord on a white wall.

To do this, it is proposed to qualify as details to exclude "non-significant" regions having an area less than a certain percentage of the total area of the image (for example 1%) and not to segment the image into regions which can cover several objects so as to avoid its sub-segmentation, which can lead to regions devoid of semantic meaning. The segmentation method according to the invention which is similar to region type techniques, operates initially on the scale of the region, starting from initial blocks of reduced size while being considered as homogeneous, so as to allow segmentation of larger objects.

The invention therefore relates to a method of spatial segmentation of an image into visual objects, characterized in that, in order to obtain objects having a semantic meaning, it comprises the following steps: - in a first phase, a partition of the image in regions according to a predetermined tiling, a fusion of neighboring regions whose similarity, according to a first similarity function, is less than a first threshold, and obtaining enlarged regions,

- in a second phase, a fusion of an enlarged region of size less than a second threshold with a neighboring region which is most similar to it according to a second similarity function.

Advantageously, the method according to the invention also comprises a third phase of fusion of the regions obtained at the end of the second phase and which are similar according to a third similarity function.

In the different phases of the segmentation method according to the invention, the similarity functions used are different in at least two of the phases.

More specifically, the subject of the invention is a method of spatial segmentation of an image into visual objects which, in order to obtain objects having a semantic meaning, comprises the following steps:

in a first phase, a partitioning of the image into initial regions according to a predetermined tiling, and a fusion of the neighboring regions whose similarity is less than a first threshold, according to a known similarity function between regions; - in a second phase, a merger of regions smaller than a second threshold with the most similar neighboring region; and

in a third phase, a representation of each of the regions obtained by a cloud of points in a representation space formed of at least one basic dimension characterizing an electromagnetic signal originating from this region and of a dimension characterizing the pixels corresponding to the values considered in the other dimensions, with

- a polynomial adjustment of a curve with the points representing each pixel of the region in the representation space of the electromagnetic signal;

- a representation of each polynomial adjustment by a series of values;

the application of a similarity function between the two sequences of values thus defined, and

- the merging of neighboring regions whose proximity between the polynomial adjustments is less than a third threshold.

In the method of the invention, the neighboring regions which can be merged in each of the phases can be initial regions, regions resulting from a fusion of initial regions, or enlarged regions resulting from previous mergers; the different types of regions thus defined can be merged with each other from the moment when the same similarity function is applied to them, and until all the similarities according to the applied function are greater than the chosen threshold. When, in the next phase, a new similarity function is applied, the merging of the regions obtained in a previous phase by application of a lower threshold similarity function is then possible until no more merging is done. possible.

The process can thus be repeated by applying a new similarity function with a higher threshold than the previous one. The regions Mergers always remain those that have not merged at the lower threshold of the previous function, including initial regions.

In addition, the similarity function between two regions can be defined by the position of the centroids of the two regions and of the ends of the two curve segments representing these two regions.

Similarly, the electromagnetic signal appearing in at least two images can be transformed in order to extract from it at least two components of movement between the two images to represent it, such as scalar value and orientation, the other process steps applying to this representation.

In the method according to the invention, the points of a cloud describing a region are distributed in a space of which three basic dimensions are three distinct linear or non-linear combinations of the three primary colors of the additive synthesis, and another dimension in this space being the number of pixels according to this distribution.

In another embodiment, the points of a cloud describing a region are distributed in a space of which three basic dimensions are the hue, the saturation and the intensity of the color and another dimension in this space being the number pixels according to this distribution.

In this case, the similarity calculation between regions is performed:

- using the hue / saturation / intensity space if an average saturation is greater than a fourth predetermined threshold;

- in the intensity space, if the saturation is less than or equal to this fourth threshold.

According to another aspect of the invention, a second segmentation on a finer scale, which may be that of the pixel, is carried out so as to obtain the precise contours of the limits of the objects, as well as their structure. internal, then a fusion of the two segmentations is carried out in order to obtain both semantically significant objects and precise contours for these objects.

By implementing the same algorithm at two different resolution scales, the results of the two segmentations of the first two steps are consistent and can be combined.

Thus, the invention also relates to a method of fine segmentation of images into semantically significant objects, consisting of:

- firstly, to extract from the image the homogeneous regions in color and texture corresponding to the objects present, neglecting objects of small size below a threshold defined as a percentage of the surface of the image, and using a representation of the regions by polynomial adjustment of their color distribution in a color space (EdC), both during the segmentation and in final representation of the segmented objects,

- to repeat the extraction with a finer resolution, in order to specify the contours, the regions obtained then corresponding to parts of objects,

in a third step, a superposition of the two preceding segmentations provides regions corresponding to the objects of the image of the first step with the precise contours of the second step, as well as an internal structure representative of these objects.

By internal structure is meant that the image is represented as a tree of objects, each higher level object being able to include one or more objects of lower level.

The last two steps can naturally be carried out at several consecutive resolutions so as to obtain a hierarchical description of the structure of composite semantic objects.

Preferably, the similarity calculation between regions can be performed: 10

- according to the method of the invention if the saturation, considered as a superposition criterion, is greater than a fourth predetermined threshold;

- in the hue-saturation-intensity space, if the saturation is less than or equal to this threshold; the predetermined thresholds can be chosen to maintain the number of regions in an interval in which the over-segmentation and the sub-segmentation do not appear and to maintain the distribution of the sizes of the regions in a defined interval to avoid over and under - segmentation.

The parameters of threshold, degree of polynomial and similarity function can be chosen adaptively by a predetermined learning method, according to thresholds of over- and sub-segmentation to be avoided, and of a calculation of predetermined evaluation of these.

In general, the similarity function applied at the end of any of the steps of the method includes the threshold parameter beyond which the fusion is not carried out.

Advantageously, composite uses of the method according to the invention are implemented:

- with the recognition of a speaker, to characterize the content of the plans and scenes of a video;

- with an automatic alignment between the script and the video, to characterize the content of the plans and scenes of a video document or to learn automatically the characteristics of the objects present in the video and in the script.

Other characteristics and uses of the invention will emerge from the following description of detailed embodiments, accompanied by the appended figures which represent respectively: 1 1

- Figure 1, an example image obtained at the end of the application of the first two broad segmentation phases of the method according to the invention;

- Figures 2a and 2b, a representation of the regions of an image in the RGB space (Red, Green, Blue) by point clouds and adjustment curves;

- Figure 3, the projections of a point cloud of RGB space on the plane (RB) and the plane (RV) to determine the ends of the adjustment interval; - Figure 4, the distribution of Euclidean distances from points to their respective lines for keyframes (characteristic images) and still images;

- Figures 5a and 5b, 6a and 6b, 7a and 7b, the segmentation into semantic objects on three images (5), (6), (7), respectively at the end of the second phase of the process (Figures 5a, 6a and 7a) and at the end of the third phase (respectively Figures 5b, 6b and 7b);

- Figures 8a and 8b, respectively a so-called wide segmentation, obtained with a resolution of 16x16 pixels, and a so-called fine resolution, of 4x4 pixels, on the same image; and - Figure 9, a superposition of the broad and fine segmentations previously obtained.

The invention is first described as an algorithm for segmenting large regions. Fine segmentation is a more particular mode that can be achieved by the same algorithm. An algorithm using two resolutions, and combining the two segmentations obtained, is described below.

According to a first exemplary embodiment, the image is first cut into a grid of so-called initial blocks, of suitable size, equal to

0.25% of the surface of the current image, that is to say: 12

- small enough so that objects of size greater than a threshold (1% in the exemplary embodiment), below which they are considered as details, are segmented, and

- wide enough to limit the computation times, which increase in a “linear” manner with the number of initial blocks.

From this breakdown, a histogram of the colors of each initial block is calculated in the RGB (Red Green Blue) space. This calculation consists in counting the number of pixels having each of the colors (of the 256 colors in the embodiment example) appearing in the image, each color being the value taken by one of the coordinates of the RGB space.

In the first fusion phase, each region is evaluated by calculating the distance between its histogram and those of the adjacent regions. The distance calculations are carried out according to the order 1 standard (L1), equal to the sum of the absolute values of the differences in the 1st degree of the values of the histograms taken two by two, the standards of order n, (Ln), being those of Minkowski (equal to the power 1 / n of the sum of the absolute values of the same differences to the power n).

The current region is merged with the one whose histogram is closest to hers, but only if the distance between their histograms is less than a high threshold which, in this embodiment, is set at 50% of the maximum distance possible from this exemplary embodiment. The higher the threshold, the higher the level of similarity of the merged regions. The merge is repeated until all the distances between adjacent regions are greater than this threshold.

At the end of this first phase, the remaining regions are either small regions, that is to say corresponding to details (size less than 1% in the embodiment), or larger and homogeneous regions, due to the high melting threshold (50% in the exemplary embodiment). These large regions can possibly be very 13

textured, that is to say with a texture granularity of the order of the size of the initial blocks,

The first phase is followed by a second phase of merging only the small regions, of size less than 1% in the embodiment example. These "details" are automatically merged with their closest neighbor by deleting the 50% fusion threshold used in the first phase, so that all the details are integrated into their surrounding region or their closest neighbor.

At this point in the algorithm, a graph of large and homogeneous regions is obtained, corresponding to large parts of objects located in the image, and possibly containing the details integrated during the second phase. Figure 1 shows the result of the segmentation at the end of the second phase. The tumulus represented T was merged despite the significant granularity of the texture of this region thanks to the use of an initial block size (1%) of the order of magnitude of the texture grain.

The entry E of the mound being smaller than the threshold of 1% in the example of embodiment, it was segmented during the first phase, then was merged during the second phase since, evaluated as detail, it could not not be considered a significant semantic object by the present algorithm.

Thus, starting from initial blocks of sufficiently large size, color distribution comparisons and not color comparisons are carried out, which makes the algorithm insensitive to textures of granularity of the order of that of the initial blocks. On the other hand, the sky C which is composed of a gradual gradient of more than 50 variations of blue is said to be over-segmented because, being a priori a semantically unique object, we had to expect a unique segmentation to form only one region. . 14

Objects, like the sky in this figure 1, can be over-segmented, the regions that compose them remaining “perceptually” similar. The third phase allows us to go beyond this stage.

The over-segmentation of the sky C illustrates the limits of the use of color histograms: they are sensitive to optical effects such as illumination, lighting variations or gradients, as described in the article by M. Stricker and M. Orengo, ^" Similarity of Color Images", Proc. SPIE'95, Storage & Retrieval for Image and Video Databases III, 1995). In addition, since the regions are homogeneous, their corresponding histograms are "hollow" (that is to say have a large proportion of values close to or equal to 0) and therefore cannot be effectively compared using the measurements of distances such as L1 or L2.

Another representation of the regions is therefore carried out in a third phase, in order to allow not only to compare the color distributions, but also the colors in the EdC (Color Space).

A representation of the regions of an image in the RGB space (Red, Green, Blue) is illustrated in Figures 2a and 2b. In FIG. 2a, each region is represented by the statistical mean of the values of its histogram forming its mean color, corresponding to the centroid C of each point cloud N1 to N6, corresponding to the color distributions of the regions, and by a polynomial interval, P, which gives an estimate of the color variations within the region.

The distributions of the colors which present a strong correlation have a curvilinear form, which justifies the estimation of these regions by a polynomial adjustment, with an interval of domain of definition of the polynomial. Figure 2b represents the point clouds corresponding to the distribution of colors appearing in the image composed of 3 15

main semantic regions R1 to R3, as well as a curve C1 (in dotted lines) corresponding to the polynomial estimate of the region R1.

At the end of the second phase, the clouds of the regions obtained are subsets, or sub-regions, of the final regions R1 to R3. In the present case, 6 sub-regions N1 to N6 have been extracted. A polynomial fit is calculated for the final regions, as well as an interval over their domain of definition. In the ideal case, the polynomial P of a final region and those Pi of the sub-regions (obtained at the end of the second phase) which compose it, are the same. And the interval of P is the union of the intervals of Pi.

The third phase of the algorithm merges regions with similar polynomial fits, and having similar, consecutive, or overlapping definition-related intervals.

In practice, to limit the number of calculations to be performed, a line is approximated using the classic method of linear regression. In the present exemplary embodiment of the algorithm, the point clouds of the regions obtained (Figure 2b) at the end of the second phase are then represented by a line segment obtained by a linear adjustment and by the centroid of the cloud N1 to N6 corresponding, which is not necessarily the center of the segment.

Each of the regions is represented by its average color and by a segment S1 to S6 on the right carried by the linear regression line of the corresponding point cloud (FIG. 2b).

For each region, as illustrated in figure 3, the point cloud of coordinates (r, v, b) of the RGB space is projected in N (r, b) on the plane (RB) as well as in N (r, v) on the plane (RV). In these plans, we respectively calculate the linear regression lines from b to r, denoted D1 in the figure, and from v to r, denoted D2. The two lines obtained are the projections on the two planes (RB) and (RV) of the line sought, denoted D3. 16

The ends E1 and E2 of the representative line segment S are the projections on the regression line D3 of the extreme points of the point cloud N. Under these conditions, the segment does not go beyond the projection of the most extreme points of the A cloud of dots. To prevent the length of the segment from being significantly influenced by the presence of colors from, for example, details, the Euclidean distance from these ends to the centroid of the cloud is preferably limited to a threshold in RGB space, threshold equal to 1.5 times the standard deviation of the color distribution for the region considered in the example embodiment.

Such an estimation of the color distributions of the regions in the EdC by a linear adjustment was tested by measuring the linear regression coefficients obtained during the adjustments, as well as the length of the segments obtained and the average Euclidean distance of the points to their respective lines. . These measurements are carried out on a collection of 4000 characteristic images (keyframes in English terminology) extracted from 4 films, and a collection of approximately 1000 still images obtained on Internet sites.

The average regression coefficients obtained are equal to 0.87 for keyframes and 0.84 for still images.

The distribution of the Euclidean distances from the points to their respective lines is shown in Figure 4. For the keyframes K, respectively the still images F, the average distance is 3.17, respectively 4.88, for an average segment length of 74.63, 90.5 respectively.

These results show a strong linear correlation between the colors present in the regions, which justifies the representation by polynomial approximation adopted in the present algorithm. 17

The regions are merged in the third phase of the algorithm by no longer comparing the color histograms, but the representative segments obtained in the second step.

The comparison of the segments is carried out in the Hue-Saturation-Intensity HSI space (initials of “Hue-Saturation-Intensity” in English terminology). This space is perceptually uniform, since providing a linear representation of the variations in the spectral frequency of a color, while the RGB space, not providing such a representation, is not suitable for such a comparison.

This comparison exploits the fact that the variations in Hue and Saturation in the regions are small, while the variations in intensity are large. The average variations observed in all the regions extracted from all the images in the previous collections are presented in Table 1. The variations in hue are expressed in degrees, the other parameters in Euclidean distance in the RGB space.

Hue Saturation Intensity

Keyframes 15.17 9.56 41.76 Pictures 20.97 18.35 48.48

TABLE 1 -

To limit, during the third step, the fusion to the perceptually similar regions, and therefore potentially belonging to the same objects of the scene, only the regions whose differences in Hue and Saturation between the centroids are less than a given threshold are merged. In the exemplary embodiment, the maximum difference in hue is fixed at a threshold equal to 7.5 °, and the maximum difference in saturation at a threshold of 15%. As shown in the example in FIG. 3, the neighboring regions R1, R2 and R3, satisfying these criteria are merged. These representative segments are close to the best polynomial adjustment, which can be obtained in the case of an adjustment of order greater than one.

When the average saturation is low, for example less than about 7% of the maximum saturation, the concept of hue is no longer representative and the segments are then only compared on the basis of their average intensity and their variation in intensity. This case corresponds to the colors "close" to the grays. Thus, the comparison method mainly consists in comparing the average gray levels (i.e. the average intensity) and the variations in intensity (i.e. the textures).

This comparison method is an approximation of a measure of similarity between polynomial fits, but is nevertheless valid since the linear correlation coefficient within the point clouds corresponding to the regions is high. The following figures show the segmentation Σ into semantic objects on three images (5), (6), (7), respectively at the end of the second phase of the process (Figures 5a, 6a and 7a) and at the end of the third phase (respectively Figures 5b, 6b and 7b).

It appears that the fusion of similar regions, obtained during the third step, provides a realistic segmentation of the images (5), (6), (7).

Another embodiment concerns the fine segmentation of objects obtained using the previous fusion process, aimed at obtaining the fine outline of these objects as well as their internal structure. The same algorithm is used with at least two different resolutions, one called fine and the other called wide. This finer resolution is obtained only by using initial blocks of smaller size, for example 4x4 pixels. The algorithm then performs the segmentation by cutting at the pixel scale, which makes it more sensitive to contours and textures.

Unlike the prior art methods mentioned above, even during fine segmentation, the areas whose surface 19

represents less than a threshold value, fixed at 1% in the example embodiment in the example embodiment, the surface of the image is not taken into account. This allows in particular to extract small objects that would not have been detected by the first segmentation due to the higher resolution. FIGS. 8a and 8b respectively present a so-called broad segmentation Σ, obtained with a resolution of 16 × 16 pixels, and a so-called fine resolution ∑f of 4x4 pixels, on the same image.

Then a superposition makes it possible to obtain an optimized final segmentation ∑o, as shown in FIG. 9; the two segmentations are obtained using the same algorithm with, as input parameter relating to the resolution, 4 X 4 pixels and 16 X 16 pixels.

The overlay retains the outlines of the fine regions included in the regions having semantic significance. However, the fine regions may not be systematically included in the corresponding wide region, because that may result from details obtained thanks to the finer resolution, for example the bars 11 on the wall at the back of the figure in figure 9.

For clarity, we use the term region to designate large semantically significant regions obtained by segmentation with large initial blocks, and the term zones to designate fine regions obtained with small blocks. The final result (Figure 9) is obtained by superimposing the two intermediate results. In the ideal case, each region would be composed exactly (to the nearest contour) of the set of zones that it spatially contains. However, three cases can arise:

• A zone is spatially included in a region and its representation is close (in the sense of the similarity measure used during the third phase of the basic algorithm) to that of this region. In this case, the area is considered to be part of the region. • 20

• Due to the more precise resolution of the fine segmentation, an area can correspond to a detail of the image that has not been extracted by the coarse segmentation (eg bars 11 on the wall behind the character of Figure 9). In this case, the distance between the representations of the zone and of the region is high, and it is considered that the zone is not part of the region, but forms a region by itself. These zones can possibly be used for the study of the spatial structure of the regions.

• A zone is not mainly included in a region (in practice, a threshold depending on the size of the zone is fixed) but extends over several regions (eg the collar 12 of the shirt of the figure in figure 9 ). In this case, the area is part of the most similar region, or is considered as a region in its own right if none of the surrounding regions is similar enough. Examples of application of the method according to the invention are described below.

Example 1: Characterization of objects; classification. All the characteristics of the segmented objects are kept for the purpose of an analysis of the document and / or an indexing of the images to allow the search for plans on their content in terms of semantic objects and actions of these objects. The set of characteristics of each region (color, texture, size, position, shape index, movement, ...) is very compact. In the case of a linear adjustment, a summary of the previous characteristics can be stored in less than 20 bytes (The position of the average color and the representative segment requiring 9). More complete representations may require a few tens of bytes per object.

An image can be represented in a form summarized by a list of descriptors of the main objects it contains, each descriptor including in particular the position and movement of the object. The 21

criteria for choosing the main objects can be for example the size, the difference in color with neighboring objects, movement, or semantic knowledge of the type "object X is important". Based on a representation of 4 objects in an image, for example, the representation of the image can be as compact as 80 bytes. To characterize an object, it is useful to know not only the descriptor of this object, but also the descriptors of neighboring objects, since an object can also be characterized by its context (ex: an airplane in the sky.). The semantics of some of the objects segmented by the proposed method can be easily extracted using their visual characteristics in a certain number of simple cases (ex: detection of day skies, lights, skin, ...) . However, the association of semantics with objects can also be based on the contribution of external knowledge. For example: "a sky is a blue or gray object with little texture, generally at the top of an image".

The problem of characterizing a semantic object is a known problem of classification or clustering of points in a multidimensional space. This classification can be done with or without learning, in supervised mode or not. In the process of the present patent, this classification is based on a compact representation of the object and, where appropriate, of the surrounding objects, in a multi-dimensional space.

Known classification methods that can be used are conventional data analysis methods, neural methods and methods using genetic algorithms.

In certain methods of data analysis, the clouds of neighboring points are characterized as clusters and projected into an adequate smaller representation space. The characterization of the cluster objects can then be done from the description by the user of a 22

or more objects in the cluster. The indexing system generalizes the characterization of one or more objects described by points of the cluster, to objects described by other points of this cluster.

In the neural approach, the system learns a "classifier", which allows dividing the representation space into clusters, from a set of examples. Examples can be provided by one of the system users during training or while in use.

In the genetic approach, there is a set of initial classifiers, each characterized by a similarity function taken from a set of possible functions, and by thresholds. These classifiers are represented by a signature which is a bit string. The initial classifiers can be drawn at random or provided by users. The user or the system determines which classifiers have given an appropriate response. The classifiers who participated in the correct answers are hybrid by recombining the signature of two of these classifiers. Random modifications of signatures or "mutations" can also be applied when creating new classifiers. For certain classes of applications, this process converges towards a population of classifiers close to the optimum. The fourth possible classification method based on the segmentation of the image into semantic objects is the search for visual objects similar to a set of given examples, based on a similarity in the characteristics of colors, shapes, etc. . The initial query obtains a global similarity function, by calculating a sum of similarity functions applied independently to different criteria, each being weighted by a value called weight. This initial request can be enriched in a known manner by allowing the user to specify which answers are satisfactory or not. A generic technique for enriching a vector query from these responses is known. In some variations of this technique, it can be based on the estimation of desirable changes in the weights of the different similarity functions, using methods derived from Bayesian probabilities. Vector research by similarity is proposed by several video indexing projects, but these projects are not based on a spatial segmentation of semantic objects and a measure of similarity according to our process.

According to the present approach, the application of these methods is facilitated by the reduced size of the descriptor, and by the possibility for the user to indicate examples and counterexamples, and if necessary to indicate whether the answer is satisfactory or not.

A script describing the content of each clip of a video is aligned with the boundaries of the clips of the video by known methods. This scπpt describes each object of a plan and its actions. By correlating the presence of an object in the script and in the video, it is possible to determine with a certain probability which object of the video corresponds to an object of the script and what are its actions. From this information, we have examples of this type of object, which allow us to automatically build a classifier for this object.

Another use of these methods in our process is to use segmentation to annotate objects by objective or subjective characteristics. To recognize the presence of one of these characteristics in an object or part of a video, it is possible to automatically choose as examples the visual objects which are annotated by this characteristic, and then proceed to learn 'one of the previous classifiers.

In the particular case where the object corresponds to a character, the recognition of the speaker by known methods of audio analysis makes it possible to choose as example several instances of the same object, and then to proceed to learn the characteristics of this objects according to one of the methods mentioned above. 24

In all cases, the segmentation and classification of the objects of a scene and their movement makes it possible to generate a script of the scene which describes the objects, the actions, and the atmosphere of the scene.

Example 2: Time segmentation of video into sequences In a known manner, a video is most often structured in shots, separated by cuts (cuts in English) or by special effects (fades, panes). A plan is a continuous series of images taken in a single take by a single camera. The segmentation of a video into plans is useful in particular for navigating the video from an interface called "story-board", which represents each plan by a characteristic image.

However, several clips can have similar semantic content, and on the other hand the number of clips in a video is often very high, which is unfavorable for effective navigation in the video from the story board.

Several authors have proposed to represent the video for a description of its content and for navigation in this content, as a sequence of sequences. According to this approach, a sequence is a series of shots describing the same environment and the same characters. The sequence is a semantic unit suitable for description of content and navigation in the video.

Methods for the automatic segmentation of a video into sequences have already been proposed. For example, a change in the evolution of the duration of the shots can characterize a change of sequences.

Another method of segmentation into sequences is based on the detection of characteristic objects. For example, a change of sequence is often linked to a change of environment, for example 25

between indoor / outdoor / day / night. The detection of an object of the day sky or night sky or lighting type optionally makes it possible to characterize a shot as shot outside day or outside night.

In the method according to the present invention, the segmentation into semantic objects and then the characterization of a certain number of objects by the methods of the preceding paragraph makes it possible to detect limits of sequences.

Two other types of time intervals are groups of shots (shots, groupings) and subjects. Plane groups have the same properties as sequences, but are not necessarily formed of contiguous planes. The subjects are a series of sequences on the same theme. Detecting subjects is particularly interesting for characterizing time intervals in documentary or news videos. The segmentation into subjects according to the present application is based on the segmentation into sequences according to the approach described above.

The detection of a subject boundary is done using one or more of the following methods:

• A known syntactic method • The presence of a semantic object belonging to a class of objects characterizing a certain subject

• The extraction of the text of dialogues and comments by known methods of ^" voice dictation ^" , the words of these texts being then classified by concepts using a method such as that previously published by Dumais and Foltz (Comm. ACM).

Once the video has been divided into subjects, the characterization of the subject by one or more descriptors is based on the same methods. Example 3: Segmentation of composite semantic objects Composite semantic objects are composed of one or more semantic regions segmented by the method described above. 26

above. In FIG. 8, the character is a composite semantic object composed of the helmet, the face, the jacket, the shirt collar, ... It is interesting to be able to find it by any region, for example by the helmet, or to visualize and annotate the complete character and not just the face.

The segmentation of such objects cannot generally be carried out solely from a still image.

It is in some cases possible, for example when a long focal length has been used for shooting, to separate the objects in the foreground which are sharp, objects in the background which are likely to be blurred . This type of segmentation cannot however be used systematically, and is likely to lead to significant sub-segmentations (example: the background segmented as a single object). Segmentation of moving objects can also be used for segmentation of composite objects in videos. Movement-based segmentation and object tracking methods have recently been the subject of numerous publications (see Proceedings of the Workshop on Image Analysis for multimedia Interactiv Services - WIAMIS'97, Louvain-la-Neuve, Belgium, June 1997). A limitation of methods based mainly on movement, compared to our approach based on semantic objects, is that the method based on movement is not always applicable (static objects) and that it does not provide information of structure. According to the present approach, it is possible to use a representation on several dimensions of the movement and of the color (for example: two dimensions for the movement, one for the number of blocks for each value of the two previous dimensions). However in another useful implementation of our process, the segmentation is done only on color, the distribution of 27

movement is calculated on each object or part of an object. This distribution (for example the mean and the standard deviation of the motion vectors) is used to define composite objects from the same motion of the different parts. The differences between the movement characteristics in different parts of a semantic object can also be used to describe a complex movement or an action of this object. For example, the movement of an arm in a character is not necessarily the average movement of the object. After compensation for camera movements (traveling shots, zooms, ...) the movement information can be used in 2 ways:

• the segmentation of moving objects returns a regional segmentation which is necessarily a superset of the segmentation obtained on still images by our algorithm. The regions belonging to the same moving object are grouped together to form a composite object;

• the regions obtained by the segmentation method presented above and having the same movement characteristics (speed, trajectory) on consecutive images are grouped together to form a composite object.

If the analysis of the movement is important for the improvement of the segmentation in composite objects, it is not sufficient in certain cases, in particular in the case of relatively static shots or scenes. A third method based on the co-occurrence of the regions in the images is proposed: if the combination of regions, for example helmet-face-jacket, appears regularly in sequences of plans, then these regions can be associated with a significant probability of co-occurrence. The objects being represented in a space of the characteristics cited above, the probability of co-occurrence of the 28

segmented semantic objects can be calculated, for example by known statistical methods used for indexing textual documents.

Finally, neighboring objects can be grouped into composite objects using their semantic value.

Example 4: Tracking an object through one or more scenes The previous methods allow you to find the same semantic object through several successive images of a scene in a video, or even through several scenes located in different passages of this video. This characterization is done using similarity methods between objects or visual regions described above, and taking into account the object's displacements: a similar object located in the same place in the following image is more likely to be the same object only if it is in an opposite part of the image. Object tracking (or tracing objects) in a video is a known problem, which is the subject of several works by other authors. In our approach, the fact of having simple or composite semantic objects limits the number of objects to follow, on the other hand we use a specific method of finding similarity between several occurrences of an object, as we have described previously.

In some cases, tracking objects can detect actions. For example, the fact that two objects move together then are separated, frequently translates the fact that one of the objects deposited the other during the interval. The fact that these are semantic objects increases the quality of this action detection.

Example 5: Selecting objects for storage on a user's system

When intelligently accessing an image base, or a large video base, or a set of broadcast audiovisual programs, it is useful to be able to store part of this base or of these 29

interesting programs for a user, on a storage system, for example at the program production site or at the user's home.

In the present method, the parts of the video stored on this storage system are time intervals or sets of images characterized by the presence of sequence descriptors or subjects verifying a request from the user or the system, or by the presence of visual or audio objects verifying such a request.

These parts are stored in a certain order, which depends either on the relevance of the objects found, or on transition rules between objects defined by the system or by the user, or in a random order. The transition rules can be extracted from the usual associations made by a user, or from the transitions usually made by the user.

The purpose of the query used is to find an object (or a sequence or a subject) in which we find with a degree of relevance and more or less high a set of content characteristics present in the query or in a game. examples associated with the query. The objects or time segments sought may be those for which either the user has expressed an interest, for example by consulting similar objects in previous sessions, or a similar user has expressed an interest.

Two users are similar compared to this application, if they show an interest in the same set of documents or audiovisual objects. In these two cases, the queries that will be used to select the objects or the time segments can be produced automatically by the system.

Example 6: Use for a compression and composition system for video objects

In a video compression system based on an object representation, it is useful to be able to describe an image as 30

a set of objects. An object that interests a user more can be transmitted with a lower compression rate than another object (such as the background). On the other hand, a video scene can be edited so as to juxtapose several objects from different scenes, or to delete certain objects.

To achieve these objectives, the arrangement of a segmentation into semantic objects is useful. In our approach, the segmentation methods used allow access to a tree structure of objects, from the time interval or the image, then from composite objects, to the internal structure of these objects, as we have previously described.

This approach makes it possible to apply the methods of a video representation system by objects, in an efficient manner, and with a granularity which varies from the composite object to the fine structure. The invention is not limited to the examples described and shown.

Other uses and applications are defined below:

- use according to which a region with semantic meaning is also characterized by smaller regions contained in the first; - to label objects in an image by a description of the type and content of these objects;

- to detect common objects in several planes of a video, with a view to grouping these planes into groups of planes with semantic meaning; - to detect sequences of shots made in the same context, called scenes or sequences, and groups of shots made in the same context, called groups of shots or clusters;

- the joint use with voice dictation to characterize the content of the plans and scenes of a video; 3 1

- application to fixed and animated images, together with annotations made by the user, to characterize the objects of the video by objective and subjective characteristics;

- application to an image base, to characterize a composite object for the frequent association of several components of the object in the same image;

- the use to describe an audiovisual object comprising images, in a representation format describing in particular the position of the semantic objects contained in the audiovisual object, these semantic objects being characterized by a set of semantic characteristics;

- the use to describe an audiovisual object comprising images, in a representation format describing in particular the actions of the semantic objects contained in the audiovisual object; the use for selecting the objects of a stream of audiovisual objects, to be stored in the storage system of a user of this audiovisual stream for the purpose of subsequent access to these objects;

- the use of the methods according to the invention, in which the degree of plausibility or achievement of the characteristic is respectively characterized or represented by a number.

- the use for compression and transmission of fixed and animated images with a variable degree of compression according to the current or previous interaction of the user with a given semantic object;

- the use for the representation of fixed and animated images as a set of semantic objects in a scene whose composition can be modified by the user.

Claims

• 32 CLAIMS

1. A method of spatial segmentation of an image into visual objects, characterized in that, in order to obtain objects having a semantic meaning, it comprises the following steps:

in a first phase, a partitioning of the image into regions according to a predetermined tiling, a fusion of neighboring regions whose similarity, according to a first similarity function, is less than a first threshold, and obtaining enlarged regions, - In a second phase, a fusion of an enlarged region of size less than a second threshold with a neighboring region which is most similar to it according to a second similarity function.

2. Method according to claim 1, characterized in that it comprises a third phase, of fusion of the regions obtained at the end of the second phase and which are similar according to a third similarity function.

3. Method according to claim 1 or 2, characterized in that the similarity functions in at least two of the phases are different.

4. Method according to one of claims 1 to 3 wherein the neighboring regions which can be merged in any one of the phases can be initial regions or regions resulting from a fusion according to a different similarity function.

5. Method according to one of claims 1 to 4, characterized in that for a similarity function a method is retained comprising: - a distribution of the pixels of a region in a representation space so as to form a cloud of points in this space, each point in this space representing a pixel there,

- this representation space comprising at least one basic dimension relating to the physical characteristics of the pixels, - an extraction of a set of statistical quantities relating to 33

this cloud, and

- the definition of a criterion of resemblance between these statistical quantities.

6. Method according to claim 4, characterized in that if the space has at least two dimensions,

- the extraction of the statistical quantities involves a polynomial adjustment of a curve with the points of the cloud of a region, and the representation of this polynomial adjustment by a series of adjusted values, and - the definition of the criterion of resemblance involves a determination a third comparison threshold for comparing such sequences of adjusted values.

7. The method of claim 6, wherein the similarity function between two regions depends on positions of cloud centroids of two regions, and of ends of two segments of curve representing these two regions.

8. The method of claim 7, wherein the similarity function is the distance between the centroids of the point clouds.

9. Method according to one of claims 6 to 8, wherein the polynomial adjustment is a linear regression.

10. Method according to one of claims 6 to 9, in which the point clouds describing a region are distributed in a space of which three basic dimensions are three ^" linear or non-linear combinations distinct from the three primary colors of additive synthesis and another dimension in this space being the number of pixels according to this distribution.

11. Method according to one of claims 5 to 9, in which the points of a cloud describing a region are distributed in a space of which three basic dimensions are the hue, the saturation and the intensity of the color, and a another dimension in this space being the number of 34

pixels according to this distribution.

12. A segmentation method according to claim 11 in which the similarity calculation between regions is carried out:

13. Method according to any one of claims 1 to 9, in which the electromagnetic signal appearing in at least two images is transformed to extract therefrom for each region at least two components of movement between the two images to represent it, such as value scalar and orientation, and where the other process steps apply to this representation.

14. A method of spatial segmentation of an image into visual objects, characterized in that: a) at least two segmentations are carried out according to one of claims 1 to 12 with at least two sizes of initial regions corresponding to different resolutions; b) the objects obtained are superimposed, c) the contour of the object obtained is replaced by the least fine resolution by the contour obtained by the finest resolution, and / or d) the image is represented as a tree structure d 'objects, each higher level object may include one or more lower level objects.

15. A segmentation method according to any one of the preceding claims in which the predetermined thresholds are chosen to maintain the distribution of the sizes of the regions within a defined interval to avoid over and under segmentation.

16. Method according to one of claims 1 to 15, in which parameters of threshold, degree of polynomial and similarity function 35

are chosen adaptively by a predetermined learning method, as a function of over-segmentation and sub-segmentation thresholds to be avoided, and of a predetermined evaluation calculation of these.

17. Use of the method according to one of claims 1 to 16, to characterize the image by a set of objects described by several characteristics, at least one of these characteristics being represented by a number or by a value in a set listed.

18. Use of the method according to one of claims 1 to 17, for classifying the visual objects of several segmented images using partition of the object representation space by a classification method with learning chosen from a method of neuronal type; a genetic type method and a data analysis method.

19. Joint use of the method according to one of claims 1 to 17 and of a measurement of the distance between the adjustment curves in a space of representation of these curves, to detect the presence of the same object in several images, consecutive or not.

20. Joint use of the method according to one of claims 1 to 17 with statistical methods for characterizing the concepts contained in a text, to characterize the content of a series of scenes relating to the same subject.

21. Joint use of the method according to one of claims 1 to 17 with the recognition of a speaker, to characterize the content of shots and scenes of a video.

22. Joint use of the method according to one of claims 1 to 17 with an automatic alignment between a script and a video, to characterize the content of the plans and scenes of a video document.

23. Joint use of the method according to one of claims 1 to 17 with an automatic alignment between a script and a video, for learn automatically the characteristics of the objects present in the video and in the script.

24. Joint use of the method according to one of claims 1 to 17 with the transcription of the speech of the video in the form of text by a voice dictation method, to characterize the content of shots and scenes of a video.