A system and method for determining the amount of an adulterant in a food product

Abstract

The invention relates to an artificial intelligence-based system and method for detecting the amount of an adulterant in liquid products such as honey, molasses and olive oil.

Description

[0001] DESCRIPTION A SYSTEM AND METHOD FOR DETERMINING THE AMOUNT OF AN ADULTERANT IN A FOOD PRODUCT

[0002] Technical Field of the Invention

[0003] The invention relates to an artificial intelligence-based system and method for detecting the amount of an adulterant in liquid products such as honey, molasses and olive oil.

[0004] State of the Art of the Invention

[0005] Adulteration is a process of adding impurities or low-quality substances to a product to increase the volume or weight thereof, or to alter the appearance or flavor thereof. An adulterant is a serious problem as it leads to health problems, economic losses and consumer fraud. There exist a number of well-known methods for detecting adulterants.

[0006] Sensory methods may be used to detect adulteration. These methods rely on the senses of sight, smell, taste and touch to detect the additive and are generally quick and easy, but they are subjective and unreliable. On the other hand, analytical methods are more accurate and subjective than the sensory methods, and they are more time-consuming and expensive. Here, impurities or foreign substances may be detected by examining the product under a microscope, or chemical tests are used to detect foreign substances or impurities.

[0007] The most commonly used adulterants are com syrup derivatives due to their high efficiency and easy accessibility. High-fiuctose com syrup (HFCS) is a sweetener made from com starch. HFCS has a chemical composition and sweetness similar to honey. When these syrups are added to honey, sensory detection of adulteration is not easy. In such cases, various analytical methods are required for the qualitative and quantitative determination of adulteration. Today, the most commonly used techniques for determining adulteration are based on chromatography and spectroscopy.

[0008] Adulteration may be detected via a resin adsorption column using high performance liquid chromatography (HPLC) with an electrochemical detector (ECD). The organic acid profile of acacia honey has been observed to be a resolution variable in the resin adsorption column. Another method of determining adulteration is based on the ratios of the stable isotopes. The nectars used by honey bees mostly come from C3 plants such as rice and sugar cane. A very small amount of nectar is collected from C4 plants such as com. Therefore, if a syrup obtained from C4 plants is used for the adulteration in honey, the use of a stable isotope ratio may be an easy separation method. The most common method used to determine stable isotopes is isotope ratio mass spectrometry (IRMS).

[0009] In addition to chromatographic separation methods, adulteration may also be determined by different spectroscopic measurements. Infrared (IR) spectroscopy is one of the most common methods used for this purpose. Fourier transform infrared spectroscopy (FTIR) is used to determine the adulteration by glucose syrup (GS), HFCS and invert syrup (IS) in the middle spectral region (4000-650 cm1). FTIR spectroscopy may successfully detect adulteration when used in combination with multivariate statistical methods. IR spectroscopy may separate C3 sugars. The IRMS method may be used for the analysis of pine honey samples adulterated with HFCS and has been observed to be a potent method in determining adulteration. Another method used to determine the carbon isotope ratios is nuclear magnetic resonance (NMR) spectroscopy. It is used to determine the adulteration with C3 plants. As an alternative to NMR spectroscopy, low-field NMR relaxometry is also used. The main disadvantage of all these mentioned methods is that they require trained personnel and are high-cost systems.

[0010] Document No. ES1257877U describes a system featuring a partially adiabatic enclosure that houses a cuvette containing the sample to be analyzed. The sample emits infrared radiation in response to temperature changes. Positioned at a distance from this setup is a thermographic camera connected to a computer running thermographic software. This system is designed to detect adulteration in food products, including honey and similar items. Although this system detects whether there is adulteration or not, it cannot provide any results regarding the amount of an adulterant. Both the presence of an adulterant and the excess thereof is a substantial problem for the consumer.

[0011] Consequently all the above-mentioned problems have made it necessary to make an innovation in the relevant field.

[0012] Objects and Summary of the Invention The main object of the invention is to detect the presence and amount of adulterant with a non- invasive method and system.

[0013] The object of the invention is to detect the presence and amount of adulterant in a quick and cost-efficient manner.

[0014] The object of the invention is to eliminate the need for qualified personnel in detecting the presence and amount of adulterant.

[0015] In order to achieve the above objects, the present invention will use a thermal camera for measuring adulteration in honey. Thermal cameras are much cheaper than chromatography, infrared and NMR spectroscopy devices, are easy to use and are an instrument with a high potential for detecting adulteration due to their working principles that do not destroy the sample during measurement. During thermal camera measurements, variables which are easy to control should be used. These are the change in the temperature range, the temperature of the environment, the environment not containing any extra infrared radiation systems and the distance between the sample and the camera. After the adulterated honey sample is heated to 60 °C, its images will be taken with a thermal camera during cooling. These images will be processed using convolutional neural networks and classified as adulterated or non-adulterated honeys, and if the honey is adulterated, the amount of the adulterant will be estimated. The equipment to be developed will not be costly and will not require an expert or trained person.

[0016] Definition of the Figures of the Invention

[0017] The figures and related descriptions used to better explain the device developed with this invention are below.

[0018] Figure 1 is a schematic representative view of the system of the invention.

[0019] Figure 2 is a flow chart of an embodiment of the inventive method.

[0020] Figure 2a is a flow chart of another embodiment of the inventive method.

[0021] Figure 2b is a flow chart of another embodiment of the inventive method, determining the input type of the adulterant. Definitions of the Elements / Features / Parts of the Invention

[0022] In order to better explain the device developed with this invention, the parts and features in the figures are numbered and the meaning of each number is given below.

[0023] 1. Adulterant detection system

[0024] 10. Casing

[0025] 20. Enclosure

[0026] 30. Thermal image element

[0027] 40. Processing unit

[0028] G. Food Product

[0029] Detailed Disclosure of the Invention

[0030] The subject of the invention relates to an artificial intelligence-based system and method for detecting the amount of an adulterant in liquid products such as honey, molasses and olive oil.

[0031] The present invention will be explained based on the detection of the presence and amount of an adulterant in honey. However, this application may also be adapted to liquid products such as molasses and olive oil.

[0032] With reference to Figure 1, the present invention comprises a casing (10). Said casing (10) has a volume in which the food product (G) to be tested for adulteration will be placed in an enclosure (20). Here, the enclosure (20) may also be provided internally to the casing (10).

[0033] At least the inner surface of the present casing (10) is provided in black, and preferably the radiation coefficient (e) is selected as 1. Thus, external thermal radiation is prevented from affecting the measurement, thereby increasing the accuracy of the measurement. Here, the black inner surface may also be provided by coating the inner surface of the casing (10) with a suitable material.

[0034] Here, a thermal camera (30) is arranged to view the interior volume of said casing (10) and wherein the thermal camera (30) takes multiple images of the food product (G) placed inside the casing (10). The thermal camera (30) may be positioned inside or outside said casing (10). The thermal camera (30) communicates with a processing unit (40) to which it transmits the images received by a processing unit (40). The communication may be a wired or wireless communication. Said processing unit (40) is configured to execute an artificial neural network trained to detect the presence and amount of the adulterant based on the images obtained from the thermal camera.

[0035] Said processing unit (40) may be a computer, or a custom processor. Preferably, said processing unit (40) is provided outside the casing (10) so as not to affect the temperature of the casing (10).

[0036] In this adulterant detection system (1), honey is firstly heated to a certain degree to perform the test. Firstly, honey is preferably heated to 60°C±2°C. Here, solutions may be used, such as keeping honey in a hot water bath to heat it. After being heated, honey is placed inside the casing (10) and left to cool to a certain temperature value, preferably 20°C, and thermal images are taken by the thermal camera (30) during this process.

[0037] Although the images may be provided at regular intervals or continuously, they are preferably taken at intervals to save processing power. In a preferred implementation, multiple images are acquired at a rate of 1 image every 10 seconds for 500 seconds. These images taken from the thermal camera (30) are fed as an input to a convolutional neural network executed by the processing unit (40).

[0038] With reference to Figure 2, said convolutional neural network includes one artificial neural network trained to detect the presence of an adulterant and another artificial neural network to detect the amount of the adulterant. Here, an artificial neural network trained to detect the presence of the adulterant is a convolutional neural network classifier (CNN classifier). If no adulteration is found in the honey tested, the process is terminated directly, but if adulteration is detected in the honey, then second artificial neural network is proceeded. Said second artificial neural network is a convolutional neural network regressor (CNN regressor). Here, the amount of the adulterant is detected, and the program is terminated.

[0039] With reference to Figure 2a, in an embodiment of the invention, the convolutional neural network regressor comprises multiple convolutional neural network sub-regressors, each of which is designed to detect the amount of one type of adulterant input. Here, after the presence of the adulterant is detected as shown in Figure 2, the amount of the adulterant may be determined by feeding the images separately to multiple convolutional neural network subregressors. Here, multiple convolutional neural network sub-regressors may be trained to detect the adulterant inputs such as fructose, maltose, glucose, and invert syrup.

[0040] With reference to Figure 2b, unlike the method described in Figure 2a, the neural network classifier is also trained to identify the input type of the adulterant (e.g., networks fructose, maltose, glucose, invert syrup). Upon determination of the adulterant type, the amount of the adulterant is detected by using the multiple convolutional neural network sub-regressors described in Figure 2a, which are trained only for the appropriate adulterant type.

[0041] In this method, the dropout layers are applied to prevent the overfitting problem while training convolutional neural network regressors, or multiple convolutional neural network subregressors. The number of the dropout layers are determined according to the formula. Wherein DL is a dropout coefficient, Iwis the width of an image (in pixel numbers), Ihis the height of an image (in pixel numbers), Dnis the number of nodes in the dense layer, and Cnis the number of nodes in the convolution layer. According to this formula, if DL is not an integer, the preceding integer is accepted as DL.

[0042] In our designed artificial intelligence model, image width = image height = 64, the number of nodes in the dense layer = 500, and as the number of convolution layers = 3, the number of dropout layers is calculated as 2.73, whereas after applying the floor function, the number of dropout layers was found to be 2.

[0043] For the training of said artificial neural network, the first artificial neural network running in the processing unit (40) resizes the image taken from the thermal camera (30) to 64x64 (Figure 2b). This neural network has a single convolution layer consisting of 32 filters in a 3x3 format. A maximum pooling layer is used to sub-sample the filter outputs. Thereafter, a flattening layer is added to flatten the two-dimensional data. A hidden layer with 100 nodes is added subsequent to the flattening layer. After the last hidden layer, the output layer has multiple outputs to classify the input image into one of the following classes. At the output, the SoftMax function is used as an activation function, and the categorical cross-entropy definition is utilized as the loss function. The RelU function is used in the hidden convolutional layer with 100 nodes. RelU is defined as RelU(x) = maximum (x, 0). As seen in Figure 2b, the first artificial neural network detects the thermal image as the adulterated or not adulterated image. If the image is adulterated, then it calculates the amount of the adulterant using the relevant artificial neural network in the second layer according to the type of the adulterant.

[0044] There are three convolutional layers in the second layer CNNs. The first layer is the convolution layer where the features of the images are obtained. 16 filters of 3x3 in size are added, and then a maximum pooling layer is added to sub-sample the filter outputs. This structure is repeated with 32 filters, and 64 filters are added in the last convolution layer. The two-dimensional data are available until this point and flattened into one-dimensional data. Upon flattening, there is a hidden layer of 500 nodes that is completely connected to the previous flattening layer. A dropout layer is added to prevent the overfitting problem. Subsequent to the dropout layer, there is another hidden layer with 100 nodes. Subsequent to the last hidden layer, another dropout layer is added. Then there is a hidden layer with 20 nodes which is connected to the single output via a sigmoid activation function. The sigmoid function defines the amount of the adulterant as a decimal number between 0 and 1.

[0045] Sigmoid(x) = 1 / (l+e^-^)

[0046] For all other activation functions, RelU is selected. The Adam algorithm is selected as an optimizer, and mean-square-error definition is used as the loss function.

Claims

CLAIMS1. A computer-aided method for detecting the presence and amount of an adulterant in liquid food products (G), characterized in that it comprises the steps of:Obtaining multiple images of a food product (G) placed in a casing (10) with at least a black inner surface during a cooling phase using a thermal camera (30),Feeding said images to a convolutional neural network classifier trained to detect whether the food product is adulterated or not, and then to a convolutional neural network regressor trained to detect the amount of an adulterant.

2. A method according to claim 1 , characterized in that said convolutional neural network classifier is trained to detect an input type of the adulterant, and said convolutional neural network regressor includes multiple convolutional neural network sub-regressor specific to the input type of the adulterant, and said images herein are fed to the convolutional neural network regressor specific to the type of the adulterant.

3. A method according to claim 1 or 2, characterized in that the dropout layers are implemented for training said convolutional neural network regressor according to the formula:and wherein DL is a dropout coefficient, Iwis the width of an image (in pixel numbers), Ihis the height of an image (in pixel numbers), Dnis the number of nodes in the dense layer, and Cnis the number of nodes in the convolution layer.

4. A method according to claim 1, characterized in that said images are obtained from a casing (10) with a radiation coefficient of 1.

5. A method according to claim 1, characterized in that said multiple images are obtained until the temperature of the food product (G) decreases from 60°C±2°C to at least 20°C.

6. A method according to claim 1 or 5, characterized in that said multiple images are obtained at intervals.

7. A method according to claim 6, characterized in that said multiple images are obtained at a rate of 1 image every 10 seconds for 500 seconds.

8. A method according to any one of the preceding claims, characterized in that said food product is honey.

9. A computer-aided system (1) for detecting the presence and amount of an adulterant in liquid food products (G), characterized in that it comprises the steps of:A casing (10) with at least black inner surface in which the food product (G) will be located in an enclosure (20),A thermal camera (30) which may be positioned so as to obtain images from inside said casing (10),A processing unit (40) to which the images obtained from said thermal camera (30) are fed, which is configured to execute a convolutional neural network classifier trained to detect whether the food product is adulterated or not, and then a convolutional neural network regressor trained to detect the amount of an adulterant.

10. An adulterant detection system (1) according to claim 9, characterized in that said casing (10) has a radiation coefficient of 1.

11. An adulterant detection system (1) according to claim 9, characterized in that said processing unit (40) is configured to execute convolutional neural network trained to detect an input type of the adulterant and said convolutional neural network regressor including multiple convolutional neural network sub-regressor specific to the input type of the adulterant, and to transmit the images to the suitable convolutional neural network sub-regressor according to the appropriate input type.

12. An adulterant detection system (1) according to claim 1 or 2, characterized in that said processing unit (40) is configured to execute a convolutional neural network regressor trained by implementing dropout layers according to the formula:and wherein DL is the dropout coefficient, Iwis the width of an image (in pixel numbers), Ihis the height of an image (in pixel numbers), Dnis the number of nodes in the dense layer, and Cnis the number of nodes in the convolution layer.

Images