ES2741073A1 - On-site detection of adulterations in food by infrared thermography and intelligent algorithms (Machine-translation by Google Translate, not legally binding) - Google Patents

Abstract

In situ detection of adulterations in food by infrared thermography and intelligent algorithms. The method of in situ detection of adulterations consists of an analysis with a thermal imager where thermographic images of pure and adulterated mixtures are obtained. The images are subsequently analyzed using mathematical models based on artificial intelligence allowing the identification of specific differences in temperature maps. Convolutional neural networks are used that are trained and optimized to distinguish qualitatively and quantitatively pure and adulterated samples with 95% medium accuracy in the case of honey and close to 100% in the case of extra virgin olive oil. This method is an analytical alternative in product quality controls and exposure to food fraud, being an innovative technique for public safety and an effective tool for the detection of hazardous compounds. (Machine-translation by Google Translate, not legally binding)

Description

DESCRIPTION

On-site detection of adulterations in food by infrared thermography and intelligent algorithms.

Technical sector

The present invention falls within the field of food quality control. More specifically, it refers to a method of detecting adulteration in food. The method combines infrared thermography and image processing by means of intelligent algorithms resulting in a simple, fast and sensitive method.

Background of the invention

Adulteration is a method of food fraud that causes a strong impact on society both in the economic and social spheres, also assuming a threat to public health. Given this situation, different methods of detecting adulteration in food have been developed.

Examples of commonly adulterated foods are honey (adulterated with syrups, cane sugar or corn), extra virgin olive oil (adulterated with worse quality oils or even expired oils) and milk (adulterated with water or mixed with sera from the dairy sector). To detect such adulterations, different authentication analytical techniques such as NMR or HPLC have been applied (Cao et al. Detection of honey adulteration with starch syrup by high performance liquid chromatography. Food Chemistry, 172 ( 2014 ), 669-674; Ribeiro et al. Detection of honey adulteration of high fruited corn syrup by Low Field Nuclear Magnetic Resonance (LF1H NMR), Journal of Food Engineering, 135 ( 2014 ), (39-43), or spectroscopy in the visible range, near infrared (Ma et al. , Qualitative and quantitative detection of honey adulterated with high-fructose corn syrup and maltose syrup by using nearinfrared spectroscopy. Food Chemistry, 218 ( 2016 ), 231-236), infrared combined with chemometrics (Ferreiro-González et al. Rapid quantification of honey adulteration by visible-near infrared spectroscopy combined with chemometrics. Talanta, 188 ( 2018 ), (288-292) or Raman (Oroian, Ropciuc, & Paduret, Floney. Adulteration Detection Using Raman Spectroscopy. Foo d Analytical Methods, 11 (4) ( 2018) , (959-968), among others. All of them, however, have limitations due to the similarity of the chemical compositions of adulterated food and adulterants, so there is still a need for new methods of detecting adulteration in situ by means of transportable, fast and sensitive equipment, capable of detection with the highest degree of reliability, safety and speed.

To give another alternative solution to this problem, the detection method of the present invention is based on the thermal properties of food by detecting variations through a thermal imager. The technology of the thermal imager has already been applied to different processes in the food industry such as product temperature control, boat filling level, solid fragment detection, packaging sealing or as a fruit control method in the agricultural sector . However, it has not been used to determine the different chemical composition of several samples of the same type of food in order to detect adulterations with a high degree of reliability and quickly and easily.

Explanation of the invention.

The present invention relates to a method of in situ detection of adulteration of food by infrared thermography using a thermal imager as the main resource and, subsequently, complementing the analysis of results with the training and modeling of non-linear algorithms by means of mathematical models centered on networks convolutional neurons as a system for identification and quantification of adulterations.

The thermal imager is able to determine the infrared radiation emitted by the sample and which is characteristic of its own chemical composition; From these infrared emissions of the samples, images are formed that are analyzed and compared by intelligent algorithms capable of detecting variations in such images and, therefore, in the samples from which they originate.

The method comprises several stages. Initially, the samples under study are heated to a temperature up to 60% higher than room temperature, to subsequently capture the thermographic images of the samples during their cooling process. With these images a database composed of thermographs of adulterated and unadulterated samples is generated. For its analysis and classification, a convolutional neural network is designed that is trained using the database for adulterated and unadulterated samples mentioned above. Mathematical models are proposed that are validated and tested for each of the foods tested. Once the mathematical models have been tested and validated, they are used to analyze images of samples suspected of adulteration, obtaining accuracy greater than 95% in both the detection and quantification of the concentration of adulterants present in the sample.

The emission of the sample does not require the use of an incident radiation on it, since the sample emits a characteristic radiation of its composition by the mere fact of being at a temperature higher than -273.15 ° C, which allows maintaining the integrity of the sample without altering its stability. Therefore, the detection is fast, simple and safe.

The camera uses infrared sensors to detect long-wave infrared emissions emitted in the electromagnetic spectrum. These emissions are related to the surface temperature of the measured sample. The result of the measurement is a color-scaled map as a function of temperature (Figure 4), visibly representing the surface temperature distribution. The thermal imager allows you to configure this color map based on its availability to distinguish the greatest visual differences in temperature. As a fundamental parameter for the control of the analysis conditions, the equipment has an internal temperature sensor responsible for measuring the ambient temperature.

The environmental temperature of the laboratory where the measurements are made is maintained between 25 ° C - 25.8 ° C. The heated samples are introduced into a partially adiabatic enclosure with a front opening that allows the radiation emitted by the sample to be captured directly by the thermographic camera (Figure 1). For this, the sample is placed at a fixed distance from the camera and videos are taken to analyze the period of temperature decrease. Then, the images are separated and analyzed, to later include them in a database for the design of intelligent algorithms.

Smart algorithms process the information provided by the thermal imager. Specifically, the mathematical model used in the present invention uses convolutional neural networks (CNN) consisting of supervised algorithms based on the extraction of image characteristics. This utility has been used in this invention for the classification of food samples and detection of adulterants if any. The function of CNNs is based on four pillars: local connections, shared weights, crushing and use of layer groups. Shared weights and local connections relate each pixel of the feature map to a neuron thus reducing calculation size and overfitting. The combination of the convolutional layers with the groupings allows to extract the set of all the characteristics from the lower to higher levels, the latter resulting from the combinations of the previous layers.

There are three different layer groups: convolutional, pooling and fully connected layers. The initial two layers form a useful tool for feature extraction while the fully connected layer is responsible for mapping previously extracted features.

Figure 2 shows a scheme of layers that make up the general CNN model together with the characteristic filters in the pooling or convolution stages.

The convolution layer is a linear operation between the matrix that makes up the (two-dimensional) images and a matrix (called a kernel filter) extractor with optimizable characteristics. This operation consists in the multiplication of the values of each position of the input matrix against the kernel filter and results in the characteristic maps for each of the images.

In convolutional networks, the most important parameters are: the size and number of filters (indicates the number of convolutional neurons that are associated with the same input region), padding (essential to maintain the dimensions of the input images, allowing the incorporation of more numbers of layers thus favoring a more exhaustive analysis of the characteristics), the stride (distance between two filter positions in the convolution). It is noteworthy that the result of the convolution requires the application of a non-linear activation function. The objective of the activation is to obtain a dispersed representation favoring the stability of the data in the face of changes and optimizing the computational work. Prior to each activation, a normalization layer is applied to accelerate the learning function of the network and limit its influence in the initial stages, thereby reducing the risk of overfitting (learning the irrelevant information of the images).

The convolutional layers work complemented with the pooling layers as a feature extractor based on the pixels of the images.

In the pooling layer the reduction of the dimensions of the input image (downsampling) is achieved, which favors the loss of non-relevant information avoiding overload of work in the networks and over-adjustment in the classifications. The result of this stage is to obtain a one-dimensional vector.

At the end of the set of convolution-poolilng layers the fully connected layer is included, responsible for the interpretation and correlation of the characteristics detected in the image. In the fully connected layer each one-dimensional vector is associated with a layer where each neuron corresponds to a class in which the image can be classified. Therefore, this layer has as its output the number of classes to which the model fits. The relationship of specific characteristics in each image is identified and adjusted to the probability corresponding to each input. For this, it is necessary to use a probability function that is introduced below.

The error function participates in the determination of the learning parameters based on the approach between the calculated class and the label corresponding to the input image. It is applied in the determination of the kernels and the weights of the established layers. The typical error function combines the softmax function (which normalizes the values obtained in the fully conntected layer for the calculation of probabilities of each class) and the cross loss. The error function is optimized by the descending gradient method, which updates the previous parameters based on the learning coefficient in the reverse direction of the error gradient.

The learning coefficient is one of the most important parameters in the initialization of the optimization of the model.

The set of images used as a database needs to be divided into three groups according to the model phase: training, validation and test images. The selection of the images that make up each of these sets is done randomly. The training group is used in the calculation of errors and optimization of learning parameters. The validation set is used for the verification of the characteristic parameters of the model and the choice of the final model. Finally, the test set certifies the final accuracy of the proposed model. To build a model that avoids overfitting, it is necessary that the images of the first two sets are completely separated from the test group. On the other hand, none of the images appear in more than one of the three sets.

Brief description of the drawings

To complement the description of the invention, figures are shown where, for illustrative and non-limiting purposes, the following has been represented:

Figure 1. Scheme of the image acquisition system formed by a partially adiabatic enclosure (2) where the cuvette is housed with the sample (1) that emits in the infrared range (3) separated a distance (4) from the thermal imager ( 5) and computer with thermographic software (6).

Figure 2. Diagram of blocks that make up the neural network.

Figure 3. Scheme of the CNN model implemented for thermographic images showing convolutional layers with ReLU (7), pooling layers (8), fully connected layers with softmax (9) and possible outputs (10).

Figure 4. Example image of a sample of honey and the color palette used in the temperature distribution of thermographic images, indicating in red the sample studied.

Figure 5. Comparison of erroneous images in the qualitative analysis of adulterated honey: (a) Acacia honey Pure straw and (b) same honey adulterated by 1%.

Figure 6. Comparison of adulterated honey samples versus pure honey samples: (a) pure lemon honey (a1) versus 8% adulterated lemon honey (a2); (b) acacia honey, honeymoon (b1) against the same adulterated honey at 2% (b2) and 8% (b3); (c) Acacia honey Pure straw (c1) against same honey adulterated by 8% (c2).

Figure 7. Comparison of erroneous images in the quantitative classification of the Acacia Pajuelo model: (a) Acacia honey Pure straw and (b) same honey adulterated by 1%.

Figure 8. Comparison of misclassified images in the combined lemon and acacia Pajuelo model: (a) pure lemon honey, (b) sample A of 1% adulterated lemon honey with rice syrup and (c) sample B of 1% adulterated lemon honey with rice syrup.

Figure 9. Comparison of samples of pure extra virgin olive oil (EVOO) and adulterated samples: (a) AVOE Picual pure; (b) adulterated with 2% sunflower oil (b1) and 8% (b2); (c) adulterated with 2% refined oil (c1) and 8% (c2); (d) adulterated with 2% pomace oil (d1) and 8% (d2)

Figure 10. Accuracy of the models for qualitative studies (yellow area) and quantitative studies (green color) for detecting adulterations in extra virgin olive oils.

Preferred Embodiment of the Invention

The present invention is illustrated by the following examples which are not intended to limit its scope.

Example 1.

This example refers to the preparation of adulterated and unadulterated food samples for later analysis. Specifically, this example refers to honey.

Three different brands are used for two monofloral honeys. Table 1 shows its commercial details and the rice syrup used as adulterant of honey.

A total of 56 adulterated samples are prepared in a range from 1% to 8% by weight. Of these samples, 8 correspond to pure honey samples (acacia and lemon), 16 to adulterated samples by 1% by weight, another 16 to adulterated samples by 4% and 16 to adulterated samples by 8%, divided by 4 brands of honey (monofloral varieties of acacia and lemon, Table 1).

Table 1. Honeys and adulterant used

TRADEMARK VARIETY ORIGIN DATE EXPIRY

Example 2

This example refers to the equipment used (Figure 1) and the method of analysis to obtain thermographic data of the analyzed samples.

An OPTRIS Pl 450 thermal imager is used with an accuracy of ± 2 ° C and a frame rate of 80Hz. Its optical resolution is collected in the dimensions of 382x288 pixels with a spectral range of 7.5-14 pm.

The distance between the thermal imager and the sample to be analyzed is kept at 25 cm. The emissivity coefficient of the honey used has been adjusted to 1. The result of the measurement is 15-minute videos where the period of temperature decrease (from 40 to 36 ° C) that corresponds, approximately, to times of 3 is analyzed. -5 minutes from which images are extracted every 15 hundredths of a second. After analyzing each image, all are included in the database for the design of the intelligent algorithms, which completes a database of approximately 80,000 images.

Thermographic images are extracted from the temperature decrease in the honey sample from 40 ° C to 36 ° C. Images are taken every fifteen hundredths of a second.

The images are classified as adulterated and unadulterated honeys. In the first group all the measured concentrations are grouped and in the second only the pure honey prepared.

A total of five convolutional neural networks are constituted, three of them referred to each of the honey brands used, one combining the two Acacia honey brands and the last one collecting all the honey brands used in the study.

Example 3

This example describes the design of a convolutional network for the treatment of thermographic results of the samples (Figure 3).

The following parameters are used for each of the layers of the convolutional networks:

- In the convolution layer, the size of filters and numbers selected for convolutional networks is 3 x 3 and 8 filters, respectively. The padding is a zero padding whose technique is characterized by filling rows and columns of zeros, maintaining the same dimensions after applying the convolution layer. The filter in the convolution is adjusted with a common jump in these networks which is 1. The result of the convolution requires the application of an activation which, in this case, is the rectified nonlinear activation function (ReLU) that calculates the function f (x) = max (0, x). Prior to each activation, the normalization layer is applied to accelerate the learning speed and limit its influence in the initial stages reducing the risk of overfitting.

- In the pooling layer, the maxpooling method is used as an ideal complement to the convolution neural network. The dimensions of the pooling are 2 x 2 with a stride (distance between two filter positions in the convolution) of 2 x2.

- The one-dimensional vector resulting from pooling is associated with the fully connected layer where each neuron corresponds to a class in which the image can be classified. To classify each kind of image the softmax function is applied:

Being o (z) j the probability value of each stipulated class, eZj the exponential of the value (j) of the input vector and e zk sum of exponentials of the values of the input vector. Each output class is associated with the own probability calculated in the mentioned function.

Table 2 shows a scheme of layers that make up the CNN model together with the characteristic filters in the pooling or convolution layers.

Table 2. Layer scheme of the convolutional model

N ° LAYER LAYER FILTER SIZE

Example 4

This example refers to the training of mathematical models.

The backpropagation algorithm used for the training of CNNs aims to reduce the differences between the estimates and the actual classes corresponding to the image database. It is based on the optimization of the convolution kernel filters and the weights in the fully connected layer through the error function and descending gradient. The descending gradient updates the previous parameters based on the learning coefficient and in the reverse direction of the error gradient following the equation:

he

w = w - a x ----- ów

where w represents the weighting weights that are the adjustable values of the model, L is the error function since it is the learning coefficient that constitutes one of the most relevant parameters in the initialization of the optimization of the model. In this case, a starts with the value 0.0001 and a rate drop of learning coefficient of 0.1. In addition, this specification uses a training data set called mini-batch, responsible for the update of the parameters, called stochastic gradient (SGD). For this case, it is fixed on all models with 64 images.

Example 5

This example refers to the validation of the models.

The images obtained as a database are divided into three groups randomly according to the model phase: training (75%), validation (15%), and test (10%). The first group is used to calculate errors and optimize the learning parameters, the second to verify the parameters of the model and the choice of the final model and the third to certify the final accuracy of the model.

Identifying adulteration models are developed in (qualitative) samples and optimized models for quantifying the concentration of adulterants in adulterated samples.

Example 6

This example describes the taking of thermographic images of pure and adulterated honey for the qualitative detection of adulterations.

The images (Figure 4) are classified as adulterated and unadulterated honeys. In the first group all honey samples with non-zero concentrations of rice syrup are grouped and in the second only pure honey prepared.

A total of five neural networks are constituted, three of them referring to each of the honey brands used, one combining the two brands of Acacia honey and the last one collecting all the honey brands used in the study (Table 3).

Table 3 shows a comparison of the amount of images dedicated in each network.

Table 3. Comparison of image base used in each model (n ° images)

NETWORK 1 NETWORK 2 NETWORK 3 NETWORK 4 NETWORK 5

CLASSES OF IMAGES Lemon Acacia Luna Acacia Mixtures Honey blends Straw floral marks

Acacia

Analyzing each network individually, it can be verified that the proposed models fit in all cases to the base of randomly selected training and validation images, which evidences the differences in the thermal behavior of the samples.

The result of the five trained networks shows an accuracy greater than 95% in all cases. It can be said that the networks allow 99% adulterated honey to be detected independently of different floral origins. In addition, the comparative result between the two brands of Acacia honey has allowed us to assert that the algorithm implemented reaches an accuracy of 99.75% in the selection of adulterated samples. These differences are justified by the influence of the treatment and the floral and climatic conditions themselves in the final composition of the honey. All these specifications originate honeys with varied characteristics perfectly analytically differentiated.

A system is made that contains all the mixtures of the presented honey variants. The convolutional network reaches a yield of 95% for this case, which implies that the structure chosen in these networks relates and classifies the variants presented between the same brands and different floral origins. The accuracy of simulated models for qualitative detection is also presented in Table 3. The average accuracy of all models reaches 99%.

Example 7

This example shows the behavior of the network in relation to the detection errors.

From the results of example 6, to better understand the behavior of the network, it is observed where the most characteristic classification errors of the models are found by comparing between the classification result of the simulation images versus the actual labeling of these images .

As a result, it can be concluded that there are certain errors in the identification of adulterations of samples with lower concentrations (1% or 2%). As shown in Figure 5, the distribution of temperatures in pure and adulterated sample at 1% has a very similar thermal behavior. For both lower concentrations, the error lies in images at temperatures of 39 ° C, which implies a fairly extended temperature distribution throughout the selected thermal range of colors. This facilitates the correlation of features in the model. The error only assumes 2.4% of the total set of added syrup concentrations between 1% and 2%, which endorses the high sensitivity of the detection method.

Example 8

This example describes the quantitative detection of syrup as adulterant of honey.

The images of each variety of monofloral honey, obtained as described in example 1, are organized according to the concentrations of adulterants and pure mixtures. Visually appreciable differences in the temperature profile with respect to the concentration of adulterant are analyzed (Figure 6). The observable differences in the thermal map can be associated with the concentration of hydroxymethylfurfural (FIMF) since this compound has been identified as an indicator of degradation of honey associated, mainly, inadequate heating in honey or adulteration with rice syrups.

As a consequence of the conditions of experimentation, CNNs focus on studying the relationship between adulterant concentrations and the differences between their characteristic maps. Thus, as shown in Figure 6 (a) and (c) for adulterations of 8%, the differences mentioned can be seen with the naked eye compared to pure honey. For concentrations as low as adulterant as 2% (Figure 6b) it is necessary to use CNNs to appreciate such differences.

CNNs focus on studying the relationship between concentrations of adulterants with the differences between their feature maps. According to Figure 6, in different types of honey the temperature distribution changes with respect to the concentration of adulterant. Even for adulterant concentrations of 8%, the differences can be seen with the naked eye compared to the pure sample. However, for lower concentrations it is necessary to use methods such as CNNs as a complement to the thermal imager to appreciate differences.

In this case, four sample classifier models are proposed depending on the concentration of adulterant. The four sample classes chosen are: pure honey, 1% adulterated honey, 4% adulterated honey and 8% adulterated honey. The result obtained for each trained CNN model is presented in Table 4.

Table 4. Trained models for quantitative identification of adulterated honey.

NETWORK 1 NETWORK 2 NETWORK 3 NETWORK 4

RED Lemon Acacia Luna Acacia Acacia

Honey Pajuelo Pajuelo +

Lemon

The values obtained in the CNNs during the training period identify adulterated samples in the previous concentrations by approximately 97%. In addition, the values obtained are verified with a batch of test images and, in this way, an accuracy of 98% is achieved.

Example 9

This example shows the behavior of the network in relation to quantization errors.

From the results of example 8, to better understand the behavior of the network, it is observed where the most characteristic quantification errors of the models are found by comparing the coincidence between the test results and the actual classes. In this case, the most notable errors observed, first, of the Acacia Pajuelo network and, secondly, the combined model have been detailed. Figure 7 shows one of the wrongly classified images. The 2% error made by the network has been mostly identified with adulterated samples of 1%. Although these errors are considered negligible, part of the difficulty of classification lies in the thermal map represented. The images corresponding to 1% wrong co-index with the color distribution of a sample at 36 ° C, which results in a simple thermal map with two colors (Figure 7), which makes it difficult to detect specific thermal characteristics at the concentration adulterant

In the combined quantification model, the identification of adulterated honey at 1%, specifically on characteristic temperature maps of 37 ° C - 36 ° C, are indicated as characteristic errors. This point coincides in detail with the Acacia model depicted in Figure 7.

For the combined model the erroneous images adulterated in samples of lemon honey are detailed (Figure 8) and it is observed that the errors are coincident with that of the individual Acacia Pajuelo model. This allows to assert that the errors are justifiable due to the high sensitivity of the models.

Example 10

The combined model is offered as an interesting alternative in quantitative analysis for a greater number of types of monofloral honeys with detection limits lower than 1% of adulterant.

This example describes the detection of adulterants in Extra Virgin Olive Oils (EVOO).

A EVOO (Picual variety) is selected and artificially adulterated with three types of worse quality oil: sunflower oil, refined olive oil and olive pomace oil (Table 5). For this, binary mixtures are prepared up to 8% by weight of adulterant. The total samples are organized equally among the three worst quality oils in concentrations of 2%, 4% and 8%. Pure and adulterated samples total a total of 24 samples.

Table 5. Oils and adulterants used

VARIETY OIL BRAND ORIGIN DATE EXPIRY

The samples are analyzed using the equipment described in example 2, taking pictures of the behavior of each mixture during a period of temperature decrease from 39 ° C to 36 ° C.

The thermal differences between the mixtures studied are due to the composition of the triacylglycerols they contain. This component is a fundamental part of oils of vegetable or animal origin and its content is a function of the origin of olive oil. In Figure 9 they show thermographic images of different samples and it can be deduced that the appreciation of differences can be invaluable by the human eye, so it requires the application of CNNs models to reach high sensitivity values.

To develop the CNN model the images are grouped individually according to the adulterant used and then optimized with the total images. Qualitative and quantitative studies are carried out. A total of eight different models are used. The training accuracy values for both studies are shown in Figure 10, where it is seen that yields close to 100% are achieved.

Claims (15)

1. Method of detection in situ of adulterations in food comprising: a) Generate a database of thermographic images of adulterated and unadulterated samples by taking said images by means of a thermal imager during a period of temperature decrease of the samples. b) Design a convolutional neural network for the treatment of the images taken. c) Train the convolutional neural network using the database to differentiate adulterated and unadulterated samples following mathematical models. d) Validate the mathematical models. e) Test the convolutional neural network designed using previously adulterated samples. f) Generate thermographic images of samples suspected of adulteration and analyze them using the previously trained and validated convolutional neural network. 2. Method of detection in situ of adulterations in foods, according to claim 1, wherein previously the samples are heated to a temperature up to 60% higher than the ambient temperature and are introduced into a partially adiabatic enclosure with a frontal opening that allows the Infrared radiation emitted upon cooling is captured directly by the terrmographic camera. 3. Method of detection in situ of adulterations in foods, according to previous claims, wherein the food is honey with possible adulteration with rice syrup in a percentage between 1 and 8% by weight. 4. Method of detection in situ of adulterations in food, according to claim 3, wherein the decrease in temperature in the sample ranges from 40 ° C to 36 ° C during which thermographic images are taken every fifteen hundredths of a second completing a base of data of approximately 80,000 images. 5. Method of detection in situ of adulterations in foods, according to claims 1 and 2, wherein the food is extra virgin olive oil (EVOO) adulterated Picual variety with three types of worse quality oil (sunflower oil, refined olive oil and / or olive pomace oil) in a percentage between 2 and 8%. 6. Method of detection in situ of adulterations in foods, according to claim 5, wherein the decrease in temperature in the sample ranges from 39 ° C to 36 ° C. 7. Method of detection in situ of adulterations in food, according to claim 1, wherein the convolutional neural network has three types of layer: convolutional, pooling layer and fully connected layer and an error function that participates in the determination of learning parameters . 8. Method of detection in situ of adulterations in food, according to claim 7, wherein the convolutional layer consists of 8 kernel filters of size 3x3, zero padding padding, a stride of 2x2 and require a non-linear activation function. 9. Method of detection in situ of adulterations in foods, according to claim 7, wherein the non-linear activation function is the rectified non-linear activation function (ReLU) and prior to it a normalization layer is applied reducing the risk of overfitting. 10. Method of detection in situ of adulterations in foods, according to claims 7 to 9, wherein the pooling layer uses the maxpooling method as a complement to the convolution with a dimension of 2x2. 11. Method of detection in situ of adulterations in foods, according to claims 7 to 10, wherein the fully connected layer applies the softmax function to rate each image. 12. Method of detection in situ of adulterations in foods, according to claims 7 to 11, wherein the training of the neural network uses a backpropagation algorithm for the optimization of the kernel filters and the weights of the fully connected layer through the function error and the descending gradient starting the learning coefficient with the value 0.0001, setting a rate drop of 0.1 and using a training data set (minibatch) called stochastic gradient (SGD). 13. Method of detection in situ of adulterations in food, according to claims 7 to 12, where the validation and testing of the models is done by randomly dividing the thermographic images obtained in three groups: training (75%), validation (15% ) and test (10%), so that the first group is used to calculate errors and optimize learning parameters 10, the second to verify model parameters and the third to check the final accuracy of the model. 14. Equipment for the detection in situ of adulterations in food, according to the claimed method, comprising a partially adiabatic enclosure (2) 15 where a cuvette is housed with the sample to be analyzed (1) that emits in the infrared range by varying its temperature (3) and that is separated at a distance (4) from a thermal imager (5) coupled to a computer with thermographic software (5). 15. Equipment for the detection in situ of adulterations in food, according to claim 14, wherein the distance between the sample to be analyzed and the thermal imager is 25 cm.