FR3113432A1 - AUTOMATIC IMAGE CLASSIFICATION PROCESS - Google Patents

Abstract

Automatic image recognition method, allowing image classification, the method comprising: a first step of reading the image by a reading cursor covering n points of the image; a measurement of intensity and gray level for each of the n points; a measurement of the intensity for each of the three colors red, green, blue at each of the n points, the sum of the values obtained for all of the n points being compared with a threshold value, the reading of the image forming four images matrix; an identification of closed polygons with non-intersecting sides; a determination of the value of a propositional function A, by an evaluation of the proportion P of polygons having straight sides, the image being classified as representing a natural landscape if the proportion P is less than a threshold value, for each of the three raster images in gray level, in green color, and in red color.

Description

AUTOMATIC IMAGE CLASSIFICATION PROCESS

The invention relates to automatic image recognition, allowing image classification.

By “image” we designate here the reproduction of a part of the universe at a given moment, this reproduction taking the form of a photograph or a video extract.

An image is presented for example in the form of a digital file in GIF, JPEG, TGA, PNG format.

By " automatic image recognition " is meant here an automatic system making it possible to classify images by type, in particular an image taken outdoors or indoors, an image including a human presence or not, an image representing an urban environment or an image representing a landscape natural.

The invention is applicable to multi-spectral digital images in which each pixel comprises several components at different wavelengths. The invention is particularly applicable to images with three RGB components.

State of the art

The image classification is carried out in the state of the art by a four-step processing.

Considering the image as a collection of regions of interest, generally of small size compared to the size of the image, these regions are detected, in a first step, by extraction of primitives.

These regions are then transformed into vectors representing the characteristics of the image, such as for example contours and/or orientations, by a second stage of description of the primitives.

A third step aims to build an image representation from descriptors generated during the second step.

The fourth step is the prediction of the presence or absence of objects in an image, using one of the supervised classification methods. Object detection in the image makes it possible to classify the image in a category, for example image including a human presence or not, urban image or landscape.

Different techniques have been implemented for each of these four steps.

The first step of extracting the primitives can be carried out by considering the entire image, by dense extraction, all the pixels of the image being processed. An example of dense extraction is presented by Jurie et al, 10th International Conf erence on Computer Vision IEEE Computer Society, 1, pp.604-610 , 2005 . This approach requires very significant time and memory resources, part of the computation time being spent processing regions that are not very informative.

In other techniques of the prior art, the extraction of the primitives is carried out by detection of the regions of interest, several families of detectors of points of interest having been proposed: detectors of contours, detectors of blobs, detectors of regions , salliance detectors.

Edge detectors detect points that correspond to a sudden change in light intensity. Examples of contour detectors are Harris detectors and their variants Harris-Laplace, Harris-affine. The points of interest are extracted along the contours taking into account only the points of maximum curvature as well as the intersections of contours.

Blob detectors detect regions that differ in their properties like color, brightness. An example of a blob detector is the Hessian detector and its Hessian-Laplace, Hessian-affine variants. A blob is generally defined as a region of the image that is brighter or darker than its surroundings. The points belonging to these regions are designated as saddle-points or col-points.

Region detectors are based on segmentation methods. An example of a region detector is the MSER ( Maximally Stable Extremal Regions ). An implementation is presented by Lee et al, Performance evaluation of local descriptors for maximally stable extremal regions, Journal of Visual Communication and Image Representation, Aug. 2017, vol.47, p. 62-72 .

Salience detectors detect the most informative regions with respect to human attention. An example is the detector resulting from the work of Laurent Itti. Intensity and color, as well as edge orientation, are often considered the most relevant visual characteristics for human attention. The identification of relevant regions in the image, for human attention, makes it possible to reduce the computational complexity.

The second stage of description of the primitives can implement differential descriptors, descriptors based on distributions, texture descriptors.

The differential descriptors are based on the calculation of derivatives of order n to approximate the neighborhood of the points of interest. An example of a differential descriptor has been proposed by Freeman et al, The design and use of steerable filters, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol 13, N°9, 1991 .

Distribution-based descriptors use histograms to represent appearance or shape characteristics of a local region of the image. The histogram describes for example the distribution of pixel intensities. Examples of such descriptors are SIFT ( Scale Invariant Feature Transform ), SURF, using Haar wavelets, HOG ( Histogram of Oriented Gradient ). An example of implementation of the SIFT, SURF, and HOG descriptors is presented in the document WO 2019106095 (Yellow Line). Reference may also be made to documents EP1850270 (Toyota), US6711293 (Univ British Col).

Texture descriptors seek to characterize the texture of an image, for example structural texture, detecting the repetition of elementary patterns, or random texture presenting a disordered aspect while appearing globally homogeneous.

The description of the texture can be performed for example by a statistical approach, using a co-occurrence or auto-correlation matrix. This approach is expensive in time and memory.

The description of the texture can be carried out by an auto-regressive model or a Markov model. The autocorrelation matrix is calculated, for a measurement of the local variations of the image at the corner points of the image, points in the vicinity of which the image varies significantly in several directions. This approach is computationally expensive.

The description of the texture can be carried out by a frequency approach. The most widely used methods are based on Fourier or Gabor transforms. The descriptors built from the Fourier transform come from images in gray levels or color images, processed plane by plane. The document US7353224 (HP) describes the use of descriptors obtained by Fourier-Mellin transform, for the identification of similar images in a database. The document FR2790571 (France Telecom) describes the use of Fourier-Mellin DF descriptors for the identification of shapes in an image, a comparison of the DF descriptors with reference descriptors being carried out after a step of learning the reference descriptors , using a neural network.

The third step aims to build an image representation from descriptors generated during the second step.

The image can be represented by a global histogram, for example color histogram, texture histogram.

The image can be represented by a histogram of statistical values calculated from the local descriptors extracted from the image, for example by bag of visual words.

The visual word bag comprises a construction of the visual vocabulary, the descriptors being extracted from an image database and then clustered by an unsupervised classification method, such as k-means, or GMM. The visual word bag then includes a construction of the coding histograms ( Vector Coding VQ, or Locality-constrained Linear C oding LLC). A variant of LLC algorithm is presented by Wang et al, A Dual Locality - Constrained Linear Coding Algorithm for Image Classification, IEEE. 6:19247-19254 2018.

The fourth step is the prediction of the presence or absence of objects in an image, using one of the supervised classification methods, in particular Bayes classifier or support vector machine. The document US7840059 (Microsoft) describes a method for recognizing an object in an image using a dictionary of textons.

Document FR3072806 (CEA) describes a method for calculating a global descriptor of an image. The global descriptor comprises a local descriptor associated with each pixel, a neighborhood being defined for each pixel associated with the local descriptor, a label being assigned to each pixel of the neighborhood, two pixels of different values having two different labels, two pixels of identical values having the same label, the set of labels in the neighborhood forming a pattern, the value of the local descriptor being assigned to a representative of an equivalence class to which the pattern belongs.

The document FR3009635 (STMicro) describes a method for searching for a similar image in an image bank from a reference image.

Object of the invention

The approaches implemented in the prior art for image classification require very large time and memory resources.

Neural network approaches require a bank of annotated images, images whose class to be predicted is known, for learning.

If the use of large memory resources and a high calculation time are acceptable for research sectors (for example artificial vision), security applications (satellite images, biometrics), medical applications, or the construction of image search engines on the Internet, the same does not apply to private, domestic or personal use.

However, the development of mobile communication terminals has transformed each user into a photographer, on a daily basis. It is estimated that the number of active smartphones is more than 7 billion, and about 50 smartphones are sold every second in the world. Every year, more than a billion photos are taken by users of smartphones, but also tablets or digital cameras. A memory card can contain several tens of thousands of photos and it is not uncommon for a personal computer user to store more than a hundred thousand photos on different media. Except for naming, or even indexing the images as they go along, it is common for a user not to be able to find an image in the mass of recorded data.

The invention aims to provide an automatic image recognition process, allowing image classification, without resorting to extensive comparison of the image with a large database.

The invention aims to provide an automatic image recognition method that can be executed by a desktop computer, or a mobile communication terminal, such as a smartphone or tablet.

For these purposes, it is proposed, according to a first aspect, to an automatic image recognition method, allowing classification of the image, the method comprising a first step of reading the image by a reading cursor represented by a matrix grid covering n points of the image, the method comprising a measurement of the intensity and of the gray level for each of the n points of the image, and a measurement of the intensity for each of the three colors red, green , blue at each of the n points of the image, the sum of the values obtained for all of the n points being compared with a threshold value, the reading of the image forming four raster images, one in gray level, one for each of the three colors red, green, blue, the method comprising a second step of identifying the closed polygons with non-intersecting sides, the method comprising, for each of the three raster images in gray level, in green color, and in red color , a step of d determination of the value of a propositional function A , the propositional function A being true when the image describes an urban landscape, the determination of the values of this function being carried out by an evaluation of the proportion of polygons having straight sides for each of the three raster images in gray level, in green color, and in red color, the image being classified as representing a natural landscape if the proportion P of polygons having straight sides is less than a threshold value, for each of the three raster images in grayscale, in green color, and in red color.

In an advantageous implementation, the threshold value of polygons having straight sides is of the order of 30%.

The method advantageously comprises a step of determining the area and the position of the centroid of the largest polygon in the raster image in blue color.

The method advantageously comprises a step of calculating the complexity of the image, the complexity of the image being calculated as a function of the number of polygons detected in the image, using the relationship:

Pi being the probability associated with the occurrence of a configuration of vertices or polygon sides, the method comprising a step of comparing the complexity value of the image with a threshold complexity value.

Advantageously, the calculation of the complexity of the image is carried out before the evaluation of the proportion of polygons presenting rectilinear sides. The fact that the result of the image complexity calculation belongs to a given value interval partly determines the descriptive characteristic of the image between artificial, urban and natural landscapes. It is then that the proportion of polygons presenting rectilinear sides is determined, the polygons which describe artificial objects being by hypothesis more regular than the polygons which describe natural objects.

Advantageously, the step of calculating the complexity of the image is carried out for the four raster images, one in gray level, one for each of the three colors red, green, blue.

In one implementation, the image is classified as representing a natural landscape if the complexity of the image falls within a threshold range of complexity values. The threshold range of complexity values is advantageously from 100 to 400.

In some implementations, the playback cursor is a square matrix grid comprising nine cells, a measurement of grayscale intensity and intensity for each of the three colors red, green, blue being made for each of the nine points of the image covered by the reading cursor, the sum of these values being compared with a threshold value, the value 1 being assigned to the points of the image in the raster image when the sum is greater than the threshold value.

Other objects and advantages of the invention will appear in the light of the description of embodiments.

Image acquisition

The method includes an image acquisition step.

Images, for example photographs in GIF, JPG, TGA, PNG formats, are saved.

In one implementation, the image is set to the dimensions 256x206 pixels in landscape format, or for portrait format with the dimensions 206x256 pixels.

Reference matrix grid

The digital data of the image, concerning the color spectrum in RGB as well as the distribution of black or white pixels in gray level, are recorded on an orthogonal reference grid.

The image is placed in an affine Euclidean vector space, equipped with an orthonormal reference. A one-to-one mapping from in ² maps to any point in the image the pair of real numbers (x _i , y _j ). The vertex of the upper left frame of the image corresponds to the point of origin of the marker.

A reference matrix grid G=(G _ij ) is constructed. The matrix G is set to the dimensions of the image, for example 256×206 pixels.

Image playback

A playback function successively scans the image in vertical slices.

The playback cursor is a frame superimposed on the image.

In one implementation, the playback cursor is represented by a square matrix grid containing 3 x 3 cells of CellLength, with nine positions indexed from 0 to 8.

The successive movements of the playback cursor on the image are incremented for each sequence by the CellLength value. CellLength is an integer of value m which can be between 1 and 9.

For each motion sequence, the value of the RGB and grayscale data present on the surface of the image are measured for the nine positions of the matrix.

A value of 0 or 1 is assigned to the nine positions, depending on the intensity recorded on the image.

A function compares the total sum of the values recorded by the cursor with an average intensity threshold, and thus determines the respective value of all the index positions i and j on the reference grid, thus producing a raster image.

The square matrix C of order n=3 and dimensions 3 x 3 with nine positions

is the playback slider, which defines a playback sequence.

Moving the playback cursor is a series of vertical translations, along the y-axis, followed by a horizontal translation along the x-axis, to cover the entire image area.

The indices i = 0 and j = 0 of the playback cursor are mapped to the coordinate point (0,0) of the mark.

A translation matrix T _y , whose domain of definition is the set of coordinate points of the image, applies a series of successive translations to the reading cursor C, following the ordinate axis.

Playback slider coefficients are incremented by the CellLength value with each sequence.

The expression of the translation matrix T _y in homogeneous coordinates is then

With Δx = 0 and Δy = m for each translation of the cursor C in the vertical direction, we obtain for the set of nine coordinate points (x, y) covered by the cursor the coordinates (x', y'):

The surface of the cursor is therefore moved vertically by the coefficient m.

When the cursor reaches the edge of the image, reduced by the coefficient m so as not to leave the limits of the image, a horizontal translation T _x is then performed on the abscissa axis.

The expression of the translation matrix T _x in homogeneous coordinates is

With Δ'x = m and Δ'y = 0 for each translation of the cursor C in the horizontal direction, we obtain for the set of nine coordinate points (x, y) covered by the cursor:

A reading sequence is therefore a vertical translation of the cursor C by the coefficient m or a horizontal translation of the cursor C by the coefficient m. That is, CellLength advances the read sequence by a length m. The value m determines the value of the increment of the displacement of the slider on the surface of the image. The value of the slider reading area is proportional to CellLength.

For each translation of coefficient m, a measurement of the intensity, 0 or 1, of the levels of gray and of the RGB color spectrum is taken on the nine points of the surface of the image covered by the matrix C of the cursor.

The sum S of the 9 values

is compared to a threshold value.

If S ≥ Threshold then [Gij]= 1. If the sum S is greater than the threshold value, then the value 1 is assigned to the indices i, j of the reference matrix grid G.

If S < Threshold then [Gij]= 0. If the sum S is less than the threshold value, then the value 0 is assigned to the indices i, j of the reference matrix grid G.

Once all the points (x, y) of the image have been read, the matrix G recorded a raster image, of resolution CellLength = m, of the image on its grid. The matrix G is a binary representation of the image, composed of 0 or 1 obtained by the successive displacement of the reading cursor C on the surface of the image. G is composed by the evaluation of the repeated sum of the values reported by the cursor C on its nine positions.

The playback cursor C can be represented by a set of nine elements, where each of the elements is associated with the value 0 or 1.

According to the binomial law, the number of possible states for the cursor C is 2 ⁹ , ie 512 possible configurations, these states all being equiprobable. These configurations can be classified into ten subsets, according to the number of elements of the cursor equal to 1. For the configurations in which the cursor C has one of its nine positions equal to 1:

9 combinations are possible, with a probability equal to P(C ₁ ) = 9/512.

Following the same reasoning, the probabilities of the other configurations for cursor C are as follows:

• C has 2 of its positions equal to 1, P(C ₂ ) = 36/512
• C has 3 of its positions equal to 1: P(C ₃ ) = 84/512
• C has 4 of its positions equal to 1: P(C ₄ ) = 126/512
• C has 5 of its positions equal to 1: P(C ₅ ) = 126/512
• C has 6 of its positions equal to 1: P(C ₆ ) = 84/512
• C has 7 of its positions equal to 1: P(C ₇ ) = 36/512
• C has 8 of its positions equal to 1: P(C ₈ ) = 9/512
• C has 9 of its positions equal to 1: P(C ₉ ) = 1/512
• C has none of its positions equal to 1: P(C ₁₀ ) = 1/512

Acquisition of tricks

The method includes a step of determining the presence of non-crossing closed polygons in the reference grids. For simplicity, in what follows, these non-intersecting polygons are called “figures”.

The perimeter of a figure is the continuous limit of the contour that draws its surface. The perimeter is a continuous observable boundary.

The acquisition of a figure is done by a continuous reading sequence along the length of its perimeter. A read sequence is the time interval required to accomplish a continuous read of the perimeter length of a figure.

A read cycle is a positive loop. The perimeters of the figures are finite, therefore, the reading sequences are finite time intervals.

Entry points on the perimeter control the start of a reading sequence. When the start point is the same as the end point, the sequence is finished and the pattern is saved.

Advantageously, the start of each read sequence is in the positive direction, in the vector plane, on the abscissa axis. The positive direction is defined by the positive incrementation of the indices i and j on the grid G. The negative direction is defined by the negative incrementation of the indices i and j on the grid G. The cursor C moves orthogonally on the grid in according to the incrementation of the indices i and j.

Each vertex implements a direction change of the playback C cursor on the perimeter of the figure.

The directions of research are four in number. The cursor moves in four directions, on two axes and in the positive or negative direction.

When the cursor is on the perimeter, it reads the nine-position square arrays that follow the edge of the figure. The edge is always parallel to the reading direction.

The description of the perimeter read during a sequence are the six positions, not empty, of the square matrix C of the cursor which is parallel to the direction of reading.

On axis I, positive direction:

On the J axis, negative direction:

On the J axis, positive direction:

On the J axis, negative direction:

We also find, on the edges of the figures, twelve possible configurations in which the boundary of the perimeter may not be rectilinear.

According to the four directions of reading, we have for a value with seven non-zero positions:

on axis I, positive direction:

on axis I, negative direction:

on the J axis, positive direction:

on the J axis, negative direction:

on axis I, positive direction:

on axis I, negative direction:

on the J axis, positive direction:

on the J axis, negative direction:

on axis I, positive direction:

on axis I, negative direction:

on the J axis, positive direction:

on the J axis, negative direction:

And for a value with 8 non-zero positions, four configurations:

on axis I, positive direction:

on axis I, negative direction:

on the J axis, positive direction:

on the J axis, negative direction:

To validate a figure, the cursor must pass through at least once in the four directions by completing a loop in the positive direction, the last vertex must be identical to the entry point.

As indicated, on the reference grid G, the figures are unintersected closed polygons.

During the sequence, the changes of direction of the reading on the perimeter are carried out on the points of vertex or vertex.

A polygon is defined by the number and direction of its vertices and the length of its perimeter, as well as the time interval needed to describe it. The smallest polygon has four vertices. The simplest non-zero area polygon is the square.

To represent the raster image of a square on a 3x3 matrix, four adjacent positions of the matrix must be non-zero.

Four configurations are possible:

The simplest polygon read in the positive direction has four vertices, each vertex being a square matrix of order n=3, with four adjacent positions equal to 1. The configurations a-d, represent the simplest possible vertex in the four reading directions.

We call respectively G00, G01, G02, G03 the square matrices, of order n=3 and of dimensions 3 x 3, with nine positions.

The matrices carry coefficients arranged on their respective positions, according to the reading direction they represent.

If we move in the matrix from coefficient 1 to coefficient 9, the reading direction represented always begins with coefficient 1.

G01 represents horizontal direction and positive direction, G03 horizontal direction and negative direction, G02 vertical direction and positive direction, G00 vertical direction and negative direction.

If we take the configuration of a) as an example we invoke

We make the sum Val of the non-zero positions on C in the case of a).

For n = 9

Val = 1 + 1 + 1 + 1 =4.

The sum W: we transfer the four non-zero positions of C to G01 and we make the sum of the corresponding values.

For n = 9

W=9+6+7+8=30.

The W value of the simplest vertex is 30.

The sum of the read square matrices denoting the edges of the perimeter and the vertices gives the value of the perimeter.

The value of the perimeter represents the number of square matrices read by the cursor C to go completely around the polygon.

The sum Val makes it possible to classify the vertices by category, according to the number of non-zero position on their matrix.

The simplest vertex with the smallest area has the value val = 4.

Using the principle that the perimeter is an observable continuous boundary and that the vertices all lie on the perimeter, the vertex with the largest area that bears the outline drawing of the figure is the nonzero eight-position vertex. So Val = 8 is the maximum value.

Below, Vertex for Val = 8, in the 4 reading directions.

If we make the sum W of the vertex which has the value 8 with the largest observable surface we obtain W = 44.

Advantageously, the reading cursor C records only the points of the perimeter which have a sum W between 30 and 44.

In other words, to be a point on the perimeter, the condition must be satisfied: 30 ≤ W ≤ 44.

To calculate the cardinal of the set V of the vertices, it is necessary to sum the cardinal of the categories of the subset of Vertex of V.

There are five categories of vertices for the Val value between 4 and 8.

To find the combinations of possible vertices, it is necessary to calculate how many possible combinations there are to fill the surface with the five remaining positions which are successively 1, 2, 3, 4 squares equal to 1. Assuming that all the draws are equiprobable , the set of possible grids that validates a vertex for each search direction is the result of this calculation.

With p representing the number of squares of value 1 and n being equal to the total number of positions on the grid,

For Val = 4, vertex with the simplest contour, there is no possibility, just 1 solution:

There is a possibility, which has a sum W such that 30 ≤ W ≤ 44.

Applying the same reasoning, the number of possibilities for which the sum W is such that 30 ≤ W ≤ 44,

• for Val = 5, five possibilities,
• for Val = 6, eight possibilities,
• for Val = 7, four possibilities,
• for Val = 8, a possibility.

The calculation of the combinations was done above for one direction, knowing that there are four directions, it is necessary to multiply by four to obtain the total number of combinations. Thus, for example, for Val=4, there is therefore a total of four possible configurations.

Image complexity calculation

The method includes, in one implementation, a calculation of the complexity of the image.

The image describes the sum of the observable discontinuities that are the polygons. The polygons are the result of the sum of the independent probabilities of the appearance of certain types of vertices or vertices as well as the sum of the independent probabilities of the appearance of the configurations of the sides at each sequence of readings on the perimeter.

The cardinality of all vertices is

Card(v) = 4 + 32 + 16 + 4 = 76.

The cardinality of all the points of the perimeter is

Card(p) = 76 + 4 +12 + 4 = 96.

The probability of obtaining any vertex is

Pv = Card(v)/ Card(p) = 76/96.

The probabilities for each vertex configuration are:

• for Val=4; P(v4) = Card(v4)/ Card(p) = 4/96=1/24
• for Val=5; P(v5) = Card(v5)/ Card(p) = 20/96 = 5/24
• for Val=6; P(v6) = Card(v6)/ Card(p) = 32/96 = 1/3
• for Val=7; P(v7) = Card(v7)/ Card(p) = 16/96 = 1/6
• for Val=8; P(v8) = Card(v8)/ Card(p) = 4/96 = 1/24

The probability of getting any side is

Pv = Card(c)/ Card(p) = 20/96 = 5/24.

The probabilities for each side configuration are:

• for Val = 6; P(c6) = Card(c6)/ Card(p) = 4/96 = 1/24
• for Val = 7; P(c7) = Card(c7)/ Card(p) = 12/96 = 1/8
• for Val = 8; P(c8) = Card(c8)/ Card(p) = 4/96 = 1/24

In one implementation, the image complexity is calculated based on the number of polygons detected in the image, using the relationship:

with Pi the probability associated with the appearance of a configuration of vertices or sides.

In one implementation, the complexity is calculated based on the total read sequence that composes the image.

with Pci the probability associated with updating the playback cursor patterns that make up the entire image.

The calculation C(I) can be appreciated as an instantaneous description and Ct(I) as an average.

Determination of polygon areas and centroid positions

The method advantageously comprises a determination of the centroids.

The figures are closed polygons of n vertices with non-intersecting sides. The simplest figures are regular polygons with four vertices, the most complex figures are irregular polygons with n vertices.

The CalculArea() function calculates the respective surface area of each polygon. The CalculCentrGrav() function calculates the respective position of the centroid of each polygon.

The method advantageously comprises a calculation of the areas of each polygon. The []SurfacePoly vector receives the respective calculation of the area of each polygon.

The polygons are numbered by the index z which follows the order of the polygons found. For the calculation we apply the Gauss formula:

The z index is numbered on the number of polygons found.

The [] PolyAX and [] PolyAY vectors contain the (x,y) coordinates of the vertices.

The index i of the vertices is numbered according to their appearance on the perimeter of the polygon, in the direction of movement of the reading cursor recorded on the vectors [] PolyAX and [] PolyAY.

In this case, vertex(x _n , y _n ) is identical to vertex(x ₀ , y ₀ ).

The method advantageously comprises a calculation of the centroids.

The vectors [] XGrav and [] YGrav receive the (x,y) coordinates of the respective centroids of each polygon.

The coordinates of the centroid of a closed polygon of n Vertex with non-intersecting sides are defined by the following formulas (to simplify the expression we set: A _z =SurfacePoly[z]):

• centroid abscissa:

• ordinate of the centroid:

Description of the image and propositional functions

The method advantageously includes a description of the image, using propositional functions.

We call I the set of all the elements contained in the image. A propositional function defined on I is an expression of the form Pr(x) which has the property of making Pr(b) true or false for each element b which belongs to I. Pr(x) becomes a statement on I, c' that is, a proposition that describes a property of I, each time an element b belonging to I is substituted for the variable x.

We call the set T _p of all the elements of I such that Pr(b) is true, the truth set of Pr(x).

We have: T _p = {x: x ∈ I, Pr(x) is true} and if {x: x ∈ I, P(x)} = I then x, Pr(x) is true; if not x, Pr(x) is false.

Similarly if {x: Pr(x)} ≠ then x such that Pr(x) is true; if not Pr(x) is false.

In an advantageous implementation, two propositional functions A and N are taken into account, the propositional functions A and N which have the domain of definition T _p , describing the result of the computation of the complexity of the image.

Depending on the number of polygons present in the set I, a sample of the polygons with the largest area is advantageously taken.

On this sample, we apply two evaluation functions, P and S, whose domain is T _p .

In one implementation, the function P evaluates the geometric characteristics of the polygons.

In one implementation, the S function calculates the large centroid distribution area based on the three colors.

Advantageously, the method implements a third propositional function W, relating to the distribution of centroids which belong to the color blue.

Advantageously, the complexity calculation and/or the calculation of the propositional functions is carried out successively

• on the elements x which belong to the set L for the light,
• on the elements y which belong to the set R for the color red,
• on the elements z which belong to the set V for the color green,

the elements x, y, and z being the results obtained for the complexity and/or the computation of the propositional functions specific to each set L, R, V.

Example of implementation

In an example of implementation, the propositional functions are built for a classification of images in two semantic categories: images of a natural landscape or image of urban structures.

x ∈ I, A(x) is true means that I is a landscape containing urban or man-made structures.

x ∈ I, N(x) is true means that I is a natural landscape containing trees, mountains or bodies of water.

The Applicant has unexpectedly found that if elements of I which are identified as non-intersecting regular or irregular polygons have straight and continuous rectilinear sides with a percentage greater than 30%, the image most often corresponds to artificial elements , especially buildings.

Conversely, the Applicant has unexpectedly found that if P has a percentage lower than 30%, the polygons are most likely natural elements of the landscape, trees, leaves, relief.

Thus, for all the elements x belonging to the set L, four situations are possible:

- the propositional function A is true;

- the propositional function N is true;

- the propositional functions A and N are both true. With the expression (A(x) N(x)), one cannot determine a result belonging only to one of the two semantic categories. The image simultaneously describes both possibilities;

- a result for propositional functions A and N cannot be determined. For the expression ¬(A(x) N(x)), there is no solution.

Similarly, for all the elements y belonging to the set R, four situations are possible:

- the propositional function A is true. Red-colored surfaces describe features of an urban landscape;

- the propositional function N is true. Red-colored surfaces depict features of a natural landscape;

- the propositional functions A and N are both true. With the expression (A(y) N(y)), one cannot determine a result belonging only to one of the two semantic categories. The image simultaneously describes both possibilities, it may be a landscape or a mixed scene;

- a result for propositional functions A and N cannot be determined. For the expression ¬(A(y) N(y)), there is no solution.

Similarly, for all the elements belonging to the set V, four situations are possible:

- the propositional function A is true. Green-colored surfaces describe features of an urban landscape;

- the propositional function N is true. Green-colored surfaces describe features of a natural landscape;

- the propositional functions A and N are both true. With the expression (A(z) N(z)), one cannot determine a result belonging only to one of the two semantic categories. The image simultaneously describes both possibilities, it may be a landscape or a mixed scene;

- a result for propositional functions A and N cannot be determined. For the expression ¬(A(z) N(z)), there is no solution.

In an advantageous implementation, the propositional function P is used to find a result for the expressions (A(x) N(x)) and ¬ (A(x) N(x)) defined on the sets L, V , R.

The applicant has unexpectedly found that if for all the variables x, y, z belonging respectively to L, V or R, the result of the calculation of the function P is greater than 30% then the function A is valid. Conversely, if the calculation of P is less than 30% then the function N is true.

In an advantageous implementation, the propositional function S makes it possible to distinguish four situations:

- situation S1, two categories of colors and light present a symmetry;

- situation S2, a category of color and the light draw a pattern;

- situation S3, light alone reveals symmetry;

- situation S4, the symmetry is absent.

In an advantageous implementation, a propositional function W, concerning the distribution of centroids which belong to the color blue, makes it possible to distinguish four situations:

- situation C1, the sky is narrow;

- situation C2, the sky is wide with a high horizon;

- situation C3, the sky is wide with a low horizon;

- situation C4, the quality of the sky is unknown.

The formula F which represents the decidability tree is expressed in its unexpanded form with the following expression:

{(A(x) N(x) (A(x) N(x)) ¬ (A(x) N(x)))

(A(y) N(y) (A(y) N(y)) ¬ (A(y) N(y)))

(A(z) N(z) (A(z) N(z)) ¬ (A(z) N(z)))}

{( S1(w) S2(w) S3(w) S4(w))}

{(C1(x) C2(x) C3(x) C4(x))}

The first part of the expression in square brackets expresses 216 possible combinations.

The inventor unexpectedly discovered that eight categories make it possible to distinguish between images of artificial landscapes and images of natural landscapes:

• artificial landscapes

(A(x) A(y) A(z))

(A(x) A(y) N(z))

(A(x) N(y) N(z))

(A(x) N(y) A(z))

• natural scenery

(N(x) N(y) N(z))

(N(x) N(y) A(z))

(N(x) A(y) A(z))

(N(x) A(y) N(z))

With a sample of 200 images chosen randomly from the categories urban landscapes or natural landscapes, we thus obtain:

83.3% success with the determination of urban landscapes.

86.2% success with the determination of natural landscapes.

The method according to the invention for the classification of images does not require significant time and memory resources.

The invention provides an automatic image recognition process, allowing image classification, without resorting to extensive comparison of the image with a large database.

The image classification method according to the invention can be executed using a desktop computer or a mobile communication terminal, for example a smartphone, a tablet.

Claims (8)

Automatic image recognition method, allowing classification of the image, the method comprising a first step of reading the image by a reading cursor represented by a matrix grid coming to cover n points of the image, the method comprising a measurement of the intensity and of the gray level for each of the n points of the image, and a measurement of the intensity for each of the three colors red, green, blue at each of the n points of the image, the sum of values obtained for all of the n points being compared with a threshold value, the reading of the image forming four raster images, one in gray level, one for each of the three colors red, green, blue, the method comprising a second step of identifying closed polygons with non-intersecting sides, the method being characterized in that it comprises, for each of the three raster images in gray level, in green color, and in red color, a step of determining the value of a propositional function A , the propositional function A being true when the image describes an urban landscape, the determination of the values of this function being carried out by an evaluation of the proportion of polygons presenting rectilinear sides for each of the three raster images in level of gray, in green color, and in red color, the image being classified as representing a natural landscape if the proportion P of polygons having straight sides is less than a threshold value, for each of the three raster images in gray level, in green color, and red color. Method according to Claim 1, characterized in that the threshold value of polygons having straight sides is 30%. Method according to claim 1 or 2, characterized in that it comprises a step of determining the area and the position of the centroid of the largest polygon in the raster image in blue color. Method according to any one of Claims 1 to 3, characterized in that it comprises a step of calculating the complexity of the image, the complexity of the image being calculated as a function of the number of polygons detected in the image , using the relationship:

with Pi the probability associated with the occurrence of a configuration of vertices or polygon sides, the method comprising a step of comparing the complexity value of the image with a threshold complexity value. Method according to Claim 4, characterized in that the step of calculating the complexity of the image is carried out for the four raster images, one in gray level, one for each of the three colors red, green, blue. 6. Method according to claim 4 or 5, characterized in that the image is classified as representing a natural landscape if the complexity of the image is included in a threshold range of complexity values. 7. Method according to claim 6, characterized in that the threshold range of complexity values is from 100 to 400. 8. A method according to any one of claims 1 to 7, characterized in that the playback cursor is a square matrix grid comprising nine cells, a measure of grayscale intensity and intensity for each of the three colors red, green, blue being carried out for each of the nine points of the image covered by the reading cursor, the sum of these values being compared with a threshold value, the value 1 being attributed to the points of the image in the raster image when the sum is greater than the threshold value.