WO2017060618A1

WO2017060618A1 - Method of classifying a product by analysing the distribution of the colours of the product

Info

Publication number: WO2017060618A1
Application number: PCT/FR2016/052556
Authority: WO
Inventors: Frédéric FOUCHER; Guillaume GUIMBRETIÈRE; Nicolas BOST; Aurélie COURTOIS
Original assignee: Centre National De La Recherche Scientifique - Cnrs -
Priority date: 2015-10-05
Filing date: 2016-10-05
Publication date: 2017-04-13
Also published as: FR3042058A1; FR3042058B1; FR3042057A1

Abstract

The present invention relates to a method of classifying a product comprising: a creation submethod (10) comprising: capture (12, 14) of a digital photograph by a detector of the product and of the template reference under identical conditions of illumination; extraction (16), from the raw signal output by the detector, of a distribution of the colours of the product and a distribution of the colours of the template reference; then classification (20) of the product with respect to a plurality of predetermined products, each predetermined product being associated with a distribution of colours and with a distribution of the colours of the template reference which were obtained by using the creation submethod, a measure of similarity being defined (22) between the distributions of colours of the product and each of the predetermined products, said measure of similarity being calculated after a calibration of the distributions of colours of the products by a calibration function (24) minimizing the difference between the distributions of the template reference which are associated with said products.

Description

METHOD FOR CLASSIFYING A PRODUCT BY ANALYZING THE COLOR DISTRIBUTION OF THE PRODUCT

DESCRIPTION

Technical area

[01] The present invention relates to the field of classification of a product among a plurality of products based on color distribution analysis.

State of the art

[02] It is known from the prior art that certain characteristics of a product can be known by colorimetric analysis.

[03] Colorimetric measurements are used in many fields. In particular :

• The agri-food sector, where color is used to control the quality of a production or the degree of maturity of fruits and vegetables for example. These checks are carried out by illuminating the products with a calibrated light source and with calibrated instruments (colorimeter). The control of a fruit in an orchard is against the eye. Note that winemakers use a refractometer to have an indication of the potential degree from the juice of a crushed grape crossed by sunlight;

• The field of printing and painting, where color is of course very important. In this area there are several levels of color measurement. To quickly find a color from a choice of paints, we use color swatches composed of the different colors available (color chart RAL K5 for example). The identification is usually in the eye but there are now applications for smartphones to find the closest color in the color chart from a photograph (ColorBlindC). Obviously, depending on the shooting conditions and the observer, the identification of the color is more or less precise. For fresco or painting restoration, the measurement of the color must be extremely precise in order to measure the pigments. This one can not be done to the eye because of the phenomenon of metamerism which makes that two colors can appear similar under a given lighting but different under another lighting. This is particularly the case for the blues. It is then necessary to use a spectrometer which breaks down the light coming from the object, and thus the color. The pigments are then chosen so as to reproduce the spectrum of the painting and not its color under a given light.

• In the field of health, we will mention the color charts used by dermatologists to define the skin at risk or those supplied with certain toothpastes to control, in the eye, the effect of the product on the whiteness of the teeth.

• Finally, in the field of scientific research, colorimetry is usually done using spectroscopic methods. It is widely used in biophysics for the study of optical contrast agents and fluorophores. Other applications use color for determination purposes such as optical microscopy on rock or ice preparations, for example.

[04] For sophisticated needs, for example the analysis of the aging of a material by the evolution of its color, one generally proceeds to a work in a controlled environment.

[05] In general, this type of analysis starts with a calibration / calibration of the shooting chain so as to obtain an absolute reference in color. For this, after setting up the lighting that will be used throughout the suite, a color chart is photographed and each camera in the chain of view (the camera, the display screen, the printer, etc.) is calibrated so that the colors reproduced at each step are as close as possible to those of the color chart. [06] Then the product to be analyzed / classified is photographed under the same lighting conditions to ensure that the colors acquired correspond to those of the reference frame.

[07] Then, the color distribution of the product is extracted, possibly standardized and serves as a basis for a classification / comparison to deduce certain physicochemical information from the product.

[08] However, this process is heavy in its implementation and requires a lot of attention in its execution. For example, if a lighting lamp fails and needs to be replaced, the entire calibration / calibration operation must be repeated.

[09] To avoid this, it was proposed to leave the colorimetric chart visible on each photograph of the product to analyze / classify so as to have on each photograph a colorimetric reference to calibrate / calibrate the photograph before doing the analysis of the product. For example, US patent application 2012/288195 describes such a solution.

[10] Nevertheless, the colorimetric analysis methods currently used have the disadvantages of requiring the knowledge of the true color of the object to be characterized. By true color, we mean its positioning in a repository known to the state of the art. This fact implies the use of a standard color chart consisting of color references positioned in the known frame according to their theoretical values. This search for the true value of colors greatly limits the classification process: Firstly by limiting the values taken by the reference colors, secondly, by the constraint of remaining in a known color reference system.

[1 1] There is therefore a real need for the development of a method for classifying a product that overcomes these defects, disadvantages and obstacles of the prior art, in particular a method making it possible to overcome the true color research, reduce costs and congestion, and to improve the classification procedure by releasing the registration constraint in a standard repository.

Description of the invention

[12] To solve one or more of the aforementioned drawbacks, a method of classifying a product to be classified comprises:

A sub-method for creating a color distribution of the product to be classified comprising:

• taking a digital photograph by a detector of the product to be classified and the reference frame under identical illumination conditions;

• extracting, from the raw signal at the output of the detector, a color distribution of the product to be classified and a color distribution of the target reference; then

Classification of the product to be classified with respect to a plurality of predetermined products, each predetermined product being associated with a color distribution and a color distribution of the reference reference, these two distributions having been obtained by using the creation sub-method of color distribution, a similarity measure being defined between the color distributions of the product to be classified and each of the predetermined products, said similarity measure being calculated after a calibration of the product color distributions by a calibration function minimizing the difference between the reference reference distributions associated with those products.

[13] Particular features or embodiments, usable alone or in combination, are:

• the reference frame is installed near the product to be classified and a single digital photo is taken of the two objects together;

• a color distribution being defined as a distribution of the number of pixels associated with each color, before the classification, a normalization of the number of pixels of the distributions is carried out; • product color distributions are transformed by a transformation function before or after calibration;

A color distribution being defined as a plurality of elementary color distributions in a predetermined repository, in particular an RGB repository, the transformation function transforms the distribution of colors into a single vector of elementary color distributions;

The similarity measure is obtained by applying a classification algorithm to the set consisting of the plurality of predetermined products plus the product to be classified, in particular an algorithm for minimizing the distances between products in a predetermined reference frame.

• The reference frame is an object whose image has a color distribution covering the color palette of the predetermined products.

[14] In a second aspect of the invention, the foregoing method is used when the products are crushed rocks, the predetermined products being rocks whose physicochemical characteristics have been measured beforehand, so that the classification of the product to classify determines its physico-chemical characteristics

[15] Another aspect of the invention relates to a computer program to be installed in a computer, comprising instructions for implementing the steps of a classification method as defined above during execution. of the program by a computing unit of said apparatus.

Brief description of the figures

[16] The invention will be better understood on reading the description which follows, given solely by way of example, and with reference to the appended figures in which:

FIG. 1 represents a histogram of the number of pixels as a function of the value of the red channel of the photo; FIG. 2 represents a flowchart of the method according to one embodiment of the invention;

FIG. 3 represents a colorimetric calibration chart;

FIG. 4 represents the structure of a film system (a), of a digital CCD detector containing a single layer of pixels arranged according to the Bayer matrix (b) and the Foveon X3 digital detector (c);

FIG. 5 represents the RGB values of the theoretical gray scale of a test pattern as a function of the measured values and the approximation by a second order polynomial shape trend curve;

FIG. 6 represents the box histograms making it possible to compare the various powders in the red, green and blue channel after correction by a 5-order polynomial and, on the right, the identification percentage for the 21 powders;

[18] Figure 7 shows the vector of the 14FR27 powder with the histograms of the R (a), G (b) and B (c) channels;

FIG. 8 represents the first 5 lines and the last line of two correspondence tables for an incorrectly identified powder (14FR24) and a correctly identified powder (14FR07); and

FIG. 9 represents the powder identification rate diagrams of the database for powders 14FR22 (a),

14FR08 (b), 14FR12 (c).

Modes of realization

[19] In preliminary, we will recall some notions useful to the understanding of the invention.

[20] The red, green, blue (RGB) system is the encoding system used by all optical imaging systems in the visible, from silver film to the eye humans, through CCD sensors or TVs. From the additive synthesis of the three primary colors it is possible to restore all the colors of less chromaticity. In computer science, every color is usually encoded in 8 bits (256 levels from 0 to 255) which allows to have 256 ³ different colors, about 16.78 million colors.

[21] A color image can be represented by a matrix of dimension MxNx3 where M and N are respectively the length and the width of the image in pixels and 3 represents the number of channels: a red, a green and a blue. The pixels are therefore identified by their coordinates (x, y, Z) or Z = R, G or B. The dynamics of the image is represented by three histograms corresponding to the distributions of the values of R, G and B on the set pixels. The maximum of a distribution is thus the value of R, G or B taken by the largest number of pixels in the image. It can be seen for example in Figure 1 that the maximum for the red channel is 88.

[22] The luminance parameter of a color in RGB is defined by: M = (R + G + B) / 3 (2.1) [23] If the RGB system is the most commonly used system, it is sometimes useful to work with other color encoding formats obtained from RGB values. The main ones are rgb, YCbCr, tsv, Yxy and L ^* a ^* b ^* . These different formats are well known to those skilled in the art.

[24] In the rest of this description, we will use the RGB format as a reference.

[25] Current cameras are fully digital and include a photosensitive detector for converting electromagnetic radiation (mostly in the visible range) into an analog electrical signal. This signal is then amplified, then digitized by an analog-digital converter and finally processed to obtain a digital image or photograph. There are three major detector families based on three different technologies: Charge-Coupled Device (CCD) technology, Complementary Metal-Oxide-Semiconductor (CMOS) technology, and Foveon technology. In the rest of this description, we will use a camera based on one of these three technologies. However, it should be noted that the choice the device and therefore the technology used by the detector has no impact on the described method.

[26] In addition, conventional cameras apply by default image processing to the raw signal recovered at the detector output, in particular to remove noise from the image or to put it in a standard digital format such as tiff, jpeg, png etc. For the purposes of the method described hereinafter, the digital image used is taken without any image processing, namely the raw image obtained at the output of the detector.

[27] In addition, we will take as an example of the use of the method the classification of volcanic rocks.

[28] With reference to FIG. 2, a method of classifying a product comprises a subprocess 10 for creating a color distribution of the product to be classified.

[29] The product is photographed, step 12, and a colorimetric chart, step 14, under identical lighting conditions.

[30] It is possible to make two successive photographs: the product then the target but, to ensure the same lighting conditions, it is often better to put the sight next to the product and take a single photograph of the two elements side beside.

[31] In the application of volcanic rocks, the rock to be studied is first crushed and then spread in a dish. A color chart is placed next to the cup and a photograph is taken.

[32] The test pattern is, for example, a colorimetric test of the market, Figure 3 as the ColorChecker ^® calibration chart composed of squares of different colors. The theoretical values R ^th , G ^th and B ^th of each square in the sRGB GMB system (standard RGB GretagMacbeth) are shown in the figure. We note in particular the bottom line whose squares represent the gray scale corresponding to an equiproportion of the chromatic coefficients. For practical reasons, these squares are numbered from 1 to 6 from white to black, and their theoretical values noted

^Th R i, G i and B ^th ^th i, with i = 1 to 6. [33] The test pattern can also be a test pattern specifically designed for the application and the products to be classified. In this case, the chart will contain a range of colors covering the color palette of the products.

[34] Then, from the (es) photo (s), one extracts, step 16, a distribution of the colors of the product to be assigned and a distribution of the colors of the reference reference.

[35] For example, as explained above, the histograms are extracted from the RGB channels.

[36] In an optional step 18, the histograms are normalized, that is to say that by applying a suitable ratio, they take into account a predetermined number of pixels: for example, 10,000 pixels. Or the high point of each histogram is positioned on the same value, for example 100.

[37] The photographed product is then to assail, step 20 with respect to a plurality of predetermined products. Each predetermined product is associated with a color distribution and a color distribution of the reference pattern, both of which distributions were obtained using the color distribution creation subprocess described above.

[38] In a particular embodiment, each product is stored as a record of a database, each record including the following information:

• An identifier that can be a name or a reference;

• a color distribution of the product;

• A color distribution of the reference pattern. These two distributions were obtained via a shooting under identical lighting condition as indicated in the steps 12, 14;

• Physicochemical characteristics associated with the product.

[39] A similarity measure is defined between the color distributions of the product to be rated and each of the predetermined products. This similarity measure is calculated after a calibration 24 of the distributions of colors of the products by a calibration function minimizing the difference between the reference pattern distributions associated with said products.

[40] Thus, for example, the treatment is as follows, for each pair produced to classify / predetermined product:

· Transformation of distributions, step 26. This optional step aims at applying a mathematical transformation of distributions to facilitate the following treatments. Note that this mathematical transformation is the same for all color distributions. Moreover, this transformation can be applied at the time of the similarity calculation or previously, for example after the normalization step. An example of a transformation is given in the example detailed below: RGB color distributions are transformed into a single vector of elementary color distributions. It is then advantageous that the records of the database contain the transformed distributions. For the continuation, the term "distribution" will be used indifferently to name the distribution or its transformed version.

• Calibration of distributions, step 24. This calibration consists of comparing the reference pattern distributions associated with the product to be classified and the product of the database and to search for a mathematical function which minimizes, if not cancels, the differences between these two distributions, then to apply this mathematical function on the distribution of the product to be classified, for example, so that the two distributions of the products share the same reference system. As described below, such a mathematical function can be a polynomial function passing through the measurement points. It should be noted that this calibration step is a relative calibration: there is no search for a "true" color as in the prior art but only a relative reference sharing of a product relative to to another. It is possible to take a product, for example the first photographed, and to label it "standard" and then to calibrate all other products in relation to this "standard" product. But even, in this particular embodiment, there is no search for "true" colors.

• Calculation of the similarity, step 22. One thus defines a measure of similarity, or a distance, in the space of the distributions of the products. Note that it is customary to define a measure of similarity as a function that believes at the same time that similarity increases and to define a distance as a function that decreases as the products come closer together. One or the other term is used interchangeably to indicate a function that varies monotonically with the similarity of the distributions.

[41] The product to be classified is then classified, step 30, by a search for the minimum of the distance, or the maximum of the similarity measure, in the predetermined reference frame.

[42] In an optional step, the physicochemical characteristics of the product to be classified are deduced from the physicochemical characteristics of the more similar product (s).

[43] It is conceivable that the implementation of the method is carried out by a computer suitably programmed by a computer program product downloadable from a communication network and / or recorded on a computer readable medium.

Example

[44] A detailed example of implementation of the above method will now be described.

[45] It relates to the classification of volcanic rocks.

[46] A camera is equipped with a Foveon X3 detector developed and supplied by Sigma ^® . This detector has an original architecture that allows it to get closer to the principle of film cameras (Figure 4). Indeed, it consists of three layers of pixels to namely, one for red, one for green and finally one for blue (Fig. 4c).

[47] At first, the images were made with a commercial Sigma SD15 digital camera. However, when shooting, this system applies different corrections to the signal. Among these, the main ones are:

• Look Up Table (LUT) linearization: the function representing the input signal as a function of the output signal is not linear. The purpose of the LUT linearization is therefore to replace this function with an affine application;

• The Dark Frame Substract: it allows to remove noise from a linearized image by subtracting from it an image of the black ("dark") itself linearized;

• Corrections are applied to remove noise or "bad pixels" that are replaced by the average value of their neighbors. These corrections may be problematic in the process because the "bad pixels" may in fact be due to a peculiarity of the powder,

• Finally, various matrix operations are then carried out in order to modify the colors on each of the three channels in order to obtain colors close to those observed by the human eye and to compensate for a brightness that is too weak or too strong.

[48] These algorithms are particularly troublesome since they will induce color changes that are difficult to control. It is therefore preferable to work on the raw signal at the output of the detector. To do this, we used a case equipped with a Foveon X3 ^® detector. It is fixed on a vertical stand to take pictures from above. Adjustable lamps are fixed on both sides of the field of view. The housing is equipped with a fixed focal length lens whose aperture of the diaphragm is adjusted manually. Finally, it is connected to a computer equipped with EPIX XCAP ^® software for processing the raw information sent by the detector and control the acquisition parameters including the exposure time (ie the time during which the detector counts the photons that happens to him), and the recording format. Images can be saved in text or tif format. All data is then processed with Matlab 2010 software.

[49] For this example, 21 powders of volcanic rocks are used. The powders are not sieved but the grains larger than 2mm are removed. The different powders are placed in a pile of random shape next to the test pattern. Different photos are taken of different piles of powders in different lighting conditions, with different exposure times and different apertures of the lens diaphragm.

[50] As the format used does not make it possible to display particularly relevant parameters, the work is done in RGB, this format being representative of the physics of the detector.

[51] The primary objective is to obtain a calibration method that makes it possible to obtain the same image from two images of the same object made under different shooting conditions. We define the calibration quality parameter, Qe:

Qe = 100 - x 100 (3.A.1)

[52] where 11 and 12 are two color images.

[53] On the test pattern, the minimum and maximum values of R, G and B are associated respectively with the black square 6 and the white square 1 (Figure 3). It seems relevant to re-spread the dynamics of the test pattern by replacing the average values of these squares with their theoretical values by using the following formulas:

_B th_ _B th

[54] R _L = x (R - R ₆ ) + R 1 ^h (3.A.5)

rth_ _r th

[55] G _L = ^G ^ - x (G - G ₆ ) + (3-A.6)

_D th_ _D th

[56] B _L = ^B ^ - x (B - B ₆ ) + B ^h (3.A.7) [57] where: [58] R1 (respectively G1 and B1) is the measured average value of the white square in channel R (respectively G and B),

[59] R6 (respectively G6 and B6) is the measured average value of the black square in the R channel (respectively G and B).

[60] With this method, the Qe factor goes from Qe = 85.71% before calibration to Qe = 95.27% after calibration.

[61] The colors of the powders forming a gray scale, the gray scale of the chart is used to calibrate the images. Figure 5 shows the theoretical values of the squares of the gray scale as a function of the average values measured on a photo made with an exposure time of 1 10 ms and an aperture of 5.6.

[62] It is observed that on no channel the response of the detector is linear. To use all the squares of the gray scale, the order 5 polynomial for connecting the average values measured for each square to the theoretical values of the target is calculated.

[63] Obtaining polynomials is done by solving the following 3 systems of equations:

25

[64] where C1, ..., C6 represents the measured average values of the squares of the gray scale, C = R, G or B. The resolution of this system made it possible to obtain the coefficients a, b, c , d, e and f in the form:

[66] with:

Y = Ci ⁴ -C ₂ ⁴

£ = Ci ² -C ₂ ²

, rth tn

2 S (C ₂ -C ₃ )

A = C ₂ ² -C | -ii≤ ^ i

v = _C -C¾¾c) ₊ (0_c _{) (C3-} _C4) , _n - rth rth ^ ( ^c 4 ~ ^c ) j_ fv μk \ _{(r 2 Γ} 2 \ ( ^{a δ θε} , * <P - ^L 4 ^L 5 ₀ + { _{λ λ0} ) ^ 4 ^) \ _{ζ οζ λζ} +

_Γ 4 _Γ 4 f (Q ³ -c |), * Γ2 ^ 1 y ^ίε _ι_ (c _Γ5 - " _{6 Λ} j

( _r5 - " _{6 Λ} j

[67] The polynomials obtained are then used to calibrate the two images of Figure 8 and calculate the Qe.

[68] A Qe of 97.53% is obtained after calibration with a polynomial of order 5. However, in order to further improve this calibration, various parameters can be taken into account:

• Acquisition in 12 bits. The images taken by the sensor are often very dark and with little contrast, therefore, when applying a correction, there are "holes" in the histograms that is to say for some values between 0 and 255 no pixels take this value. To avoid this problem, the images are acquired in 12 bits instead of 8 bits then the method of the order 5 polynomials is used to obtain a calibrated 8-bit image.

• Validity limits. It is obvious that calibration will have limits: a photograph taken in the dark or strongly overexposed can never be corrected. [69] Figure 6 shows the intervals created by the maximum distributions in the red, green and blue channels, obtained from 6 12-bit photos calibrated with the 5-order polynomial method for the different powders, after verification of their luminance level. It is observed that the overlap of the intervals is relatively limited. It thus seems possible to hope to use the images made for identification purposes. On the other hand, even if it is impossible to differentiate one powder from all others, a link between color and chemical composition could explain the overlap of intervals.

[70] In order to obtain an identification method, a database is created from the images of the corrected powders used as references for the analysis of unknown powders. In order to take into account, for the same powder, variations after the correction as well as the three channels simultaneously, the histograms are extracted from the various powders after correction. For each image a vector 1 x768 is created and contains in [0,255] the values of the histogram in the red channel, in [256,512] the values of the histogram in the green channel and in [513,768] the values of the histogram in the blue channel. To form the database, for each powder the sum of the vectors is obtained from the 6 normalized images by the distribution maxima in each channel in order to overcome the number of pixels used. Thus, if for one powder one channel is predominant over another, it is possible to keep this information. We can see in Figure 7 that there are more pixels with the same value in the green channel than in red or blue.

[71] Once these vectors were created for each of the 21 powders, the database was tested. For this, photos of the powders were taken under different conditions (opening of the diaphragm, exposure time) and for different form of pile. The images have been corrected and the vector associated with each calculated image. In order to compare the obtained vectors with those of the database one applies the following approach: • be v the vector of the powder to be identified, vi i <≡ D 1, 21 □ the powders vectors of the database, n = --- with n _p the number of pixels of the powder and n _r , n _g , nb the distribution maximums previously used for the normalization of the powder to be identified and n _t = - with n _p , the number of pixels of the powder i and n _r i, n _ri + n _gi + n _bi

n _g i, ribi the coefficients previously used for the normalization of the powder i, ieD 1, 21 □;

• This gives a percentage of correspondence pi given by:

\ v- V; \

Pi = 100 - x 100

n + Ui

• For a given image, one calculates the percentage of correspondence of its vector with those contained in the base then one sorts by decreasing value. This gives a list with the most probable rock first (Figure 8).

[72] Using the 197 images made on the 21 powders, the tests carried out give the good powder as being the one with the best match rate in 84% of the cases. In addition, when the good powder does not appear first, the rock type remains correct as shown in FIG. 8 in the case of the 14FR24 powder. In the end, the results are more than conclusive.

[73] In order to ensure that the method used clearly demonstrates a link between the color of the rocks and their composition, a powder from the database has been removed. The 8 images of this powder were taken, calibrated and tested with the database of other 20 powders. The powders with the five highest identification percentages were found. Using this method on three randomly chosen powders: 14FR22 (Fig. 9a), 14FR08 (Fig. 9b) and 14FR12 (Fig. 9c). FIG. 9 represents the powders identified during the tests with the corresponding percentage. We can thus notice that for the 14FR22 powder is the 14FR02 powder that seems the closest. However, these two powders have a similar percentage by weight of Na 2 O + K 2 O. For the 14FR08 powder, it may be noted that it is the 14FR1 1 powder which is the closest. And these two powders have the same silica content.

[74] The invention has been illustrated and described in detail in the drawings and the foregoing description. This must be considered as illustrative and given by way of example and not as limiting the invention to this description alone. Many alternative embodiments are possible.

[75] The method described has wider fields of application than the classification of rocks which served as a support example above. Indeed, with the explosion of the smartphone market in recent years, digital cameras have become tools used daily. The process can therefore find many applications for the general public as for scientific research through industry, agri-food or medical research for example. Some simple examples of consumer applications can be mentioned:

• Measuring the degree of maturity of fruits and vegetables. For example, it would be possible to know if a melon or a peach is ripe enough on the stall of a primeur without touching it or cutting it;

• Control of the freshness of a product. For example, the condition of a meat, if the cold chain has been broken;

• Monitoring the alteration of facades of historical monuments. These monuments suffer alterations related to bad weather and pollution and the photographic follow-up may make it possible to envisage restorations and repairs before the stones are at an advanced stage of deterioration;

• The identification of narcotic or explosive products using a smartphone or drone;

[76] Until now, to achieve accurate color measurements, it is necessary to work with calibrated instruments and lighting calibrated. This type of measurement is therefore difficult on site and may require the use of expensive instruments (spectrometers). The method makes it possible to dispense with the light source. It can therefore be used on site, including in places inaccessible to humans (on Mars for example). It will be usable with a simple digital camera, including a mobile phone. The purchase of any dedicated device is necessary which allows access to the method to the general public. The use of a simple digital image sensor makes the ultra-portable analysis method compatible with drones, for example.

[77] In the claims, the word "comprising" does not exclude other elements and the indefinite article "a" does not exclude a plurality.

Claims

A method of classifying a product to be classified comprising:

A subprocess for creating (10) a color distribution of the product to be classified comprising:

Taking a digital photograph (12, 14) by a detector of the product to be classified and of the reference frame under identical illumination conditions;

• extraction (16), from the raw signal output of the detector, a color distribution of the product to be classified and a color distribution of the reference reference; then

Classifying (20) the product to be classified with respect to a plurality of predetermined products, each predetermined product being associated with a color distribution and a color distribution of the reference frame, both of these distributions having been obtained by using the sub-product; a method of color distribution creation, a similarity measure being defined (22) between the color distributions of the product to be classified and each of the predetermined products, said similarity measure being calculated after a calibration (24) of the product color distributions by a calibration function minimizing the difference between the distributions of the reference reference associated with said products.

Method according to claim 1, characterized in that the target reference is installed near the product to be classified and a single digital image is taken of the two objects together.

Method according to claim 1 or 2, characterized in that, a color distribution being defined as a distribution of the number of pixels associated with each color, before the classification, a normalization of the number of pixels of the distributions is carried out. Method according to claim 1, 2 or 3, characterized in that the product color distributions are transformed (26) by a transformation function before or after the calibration.

A method according to claim 4, characterized in that, a color distribution being defined as a plurality of elementary color distributions in a predetermined repository, in particular an RGB referential, the transform function transforms the color distribution into a single vector of distributions elementary colors.

Method according to Claim 5, characterized in that the similarity measure is obtained by applying a classification algorithm to the set consisting of the plurality of predetermined products plus the product to be classified, in particular an algorithm for minimizing distances. between products in a predetermined repository.

Method according to one of the preceding claims, characterized in that the target reference is an object whose image has a color distribution covering the color palette of the predetermined products. Use of the method according to any one of the preceding claims, characterized in that the products are crushed rocks, the predetermined products being rocks whose physicochemical characteristics have been measured beforehand, so that the classification of the product to be classified determines its physicochemical characteristics. Computer program product downloadable from a communication network and / or recorded on a computer readable medium and / or executable by a processor, characterized in that it comprises program code instructions for implementing the method according to at least one of claims 1 to 7.