A method for classifying an image into a set of predefined classes, comprising a computer program and an associated electronic classification system.

A classification method using a learning model with twice the same first neural network and a second neural network addresses adversarial examples and out-of-domain detections, improving robustness and simplifying implementation in resource-constrained systems.

FR3141787B1Active Publication Date: 2024-11-01THALES SA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
FR2022011668
Authority / Receiving Office
FR · FR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2022-11-09
Publication Date
2024-11-01
Estimated Expiration
2042-11-09

AI Technical Summary

Technical Problem

Existing classification models for images are vulnerable to adversarial examples and out-of-domain detections, leading to false alarms and action protocols, and auxiliary verification mechanisms complicate implementation in resource-constrained systems.

Method used

A classification method using a robust classification model trained with a learning model comprising twice the same first neural network and a second neural network, with a predefined combination function, to reduce erroneous classifications.

Benefits of technology

The method enhances classification robustness by reducing false alarms and action protocols, while being suitable for resource-constrained systems by simplifying implementation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000022_0000
    Figure 00000022_0000
  • Figure 00000023_0000
    Figure 00000023_0000
  • Figure 00000024_0000
    Figure 00000024_0000
Patent Text Reader

Abstract

Method for classifying an image to be classified into a set of predefined classes, set of associated computer program(s) and electronic classification system. The present invention relates to a method for classifying an image (12) into a set of predefined classes, based on a classification model (110). The method comprises a step of applying a machine learning algorithm to a learning model trained from a first and a second neural network to obtain a first trained neural network and a second trained neural network. The method further comprises the generation of the classification model (110) by connecting the trained neural networks. The method includes an exploitation phase to determine a class associated with an image to be classified based on the classification model. The learning model comprises two identical first neural networks.The classification model includes a single first trained neural network. Figure for the abbreviation: Figure 1.
Need to check novelty before this filing date? Find Prior Art

Description

Title of the invention: Method for classifying an image to be classified among a set of predefined classes, set of computer program product(s) and associated electronic classification system

[0001] The present invention relates to a method for classifying an image to be classified among a set of predefined classes, from a classification model, the method being implemented by an electronic classification system.

[0002] In the field of surveillance, it is known to use models connected to an image sensor to detect the presence of aircraft such as airplanes, helicopters or drones near a surveillance area.

[0003] For example, the detection of drones around an airport is a current issue because these drones can disrupt aircraft taking off and landing at the airport and thus disrupt the entire operation of the airport.

[0004] Similarly, it is known to monitor aircraft flying over a protected area or a dangerous area, for security reasons and / or strategic reasons.

[0005] For this purpose, the use of a neural network classification model is known for classifying images. Classification models are known in particular, comprising two interconnected neural networks, the first of which is an encoder and the second of which is a classifier. When an image is supplied to the encoder, the latter is intended to produce a vector comprising a reduced number of digital components, representative of the content of the image. The second neural network is a classifier intended to classify whether the vector representing the content of the image corresponds to one of a plurality of predefined classes. The predefined classes make it possible, for example, to distinguish whether an object represented in the image is authorized or prohibited.

[0006] These classification models are then generally connected to the input of an alarm system capable of emitting an alarm and / or triggering an action protocol to be implemented, when a prohibited object is detected by the classification model.

[0007] However, such models suffer from two major problems that make them not very robust.

[0008] The first problem concerns "adversarial examples." These adversarial examples are images that represent content associated with a predefined class but have been slightly modified, in a way that is imperceptible to the human eye, in order to trick the classification model into classifying it in another of the predefined classes. For example, an image representing a drone can be modified so that the classification model considers it to be a helicopter without this modification being visible to the human eye.

[0009] Thus, during an attack on the classification model, this type of modification of the images to be classified is discreet, since it is invisible to the human eye.

[0010] The second problem concerns so-called "out of domain" or OOD detections. These detections concern the classification, in one of the predefined classes, of an object that does not correspond to any of the classes. For example, if a particular classification model is intended to classify images in one of the following three predefined classes: airplane, helicopter and drone, it happens that an image representing a raptor is considered by the classification model as representing an airplane. Thus, although the image represents a raptor and not an airplane or any motorized vehicle, the classification model will classify this image in the class: airplane.

[0011] The two aforementioned problems are therefore likely to lead to false alarms and / or action protocols triggered unnecessarily.

[0012] To make the operation of classification models more robust, auxiliary verification mechanisms are known. These mechanisms include auxiliary neural networks that verify the classification performed by the classification model. However, these mechanisms make the implementation of the model more cumbersome and hinder the embedding of the model. Indeed, if a classification model is intended to be embedded in a mobile system for which memory and / or computing power is limited, it is generally complex, or even impossible, to also embed auxiliary neural networks responsible for verifying the classification.

[0013] The present invention therefore aims to propose a classification method which uses a robust classification model and which can be embedded.

[0014] To this end, the present invention relates to a method for classifying an image to be classified among a set of predefined classes, from a classification model, the method being implemented by an electronic classification system,

[0015] the method comprising a learning phase comprising the following steps: - training a learning model from a first neural network and a second neural network, - obtaining a training data set from the training model, - applying a machine learning algorithm to the training model, from the training data, to obtain a first neural network trained from the first neural network, and a second neural network trained from the second neural network, - generation of the classification model by connection of the trained neural networks,

[0016] the method further comprising an operating phase comprising the following steps: - reception of the image to be classified, - determination of the class associated with the image to be classified by providing the image to be classified as input to the classification model,

[0017] wherein during the training step, the trained learning model comprises twice the same first neural network, a predefined combination function connected to the output of the first neural networks, and the second neural network connected to the output of the predefined combination function, each training data comprising a pair of distinct images each for a first neural network,

[0018] during the step of applying the machine learning algorithm, the first neural networks of the learning model being identical to each other, and

[0019] during the generation step, the classification model comprising a single first trained neural network.

[0020] By training the learning model comprising the first neural network twice, and the learning algorithm applied to this learning model, the first and second neural networks are trained in a particular manner which intrinsically allows the number of erroneous classifications to be reduced.

[0021] According to other advantageous aspects of the invention, the classification method comprises one or more of the following characteristics, taken in isolation or in all technically possible combinations: - the image to be classified represents a drone, an airplane or a helicopter,

[0022] the set of classes preferably comprising at least one of the following classes: • a drone, • an airplane, and • a helicopter; • the first trained neural network is an encoder capable of converting an image into an output vector representative of the content of the image,

[0023] the second trained neural network being a classifier capable of determining, from an output vector representative of the content of an image, a class to which the content of the image belongs among the set of predefined classes; - each training data further includes a class vector associated with the pair of images;

[0024] during the application step, the learning algorithm being such that, following its application, the learning model is capable of providing a vector substantially identical to the class vector when the image pair is provided as input to said learning model; - the class vector comprises a plurality of components each associated with a class such as: • if the two images belong to the same class, the component associated with said class has a first predefined value preferably equal to 1, and all the other components have a second predefined value preferably equal to 0, and • otherwise, all the components are identical to each other, preferably equal to the ratio between 1 and the number of classes; - each first and second neural network comprises several layers of neurons and adjustable synaptic weights respectively connecting two neurons from two successive layers,

[0025] the step of applying the automatic learning algorithm comprising the adjustment of the adjustable synaptic weights of the first and second neural networks such that, for each training data item, the supply of each image to a respective first neural network makes it possible to obtain at the output of the training model, a vector substantially identical to the class vector of the training data item; - each first neural network is capable of providing, at its output, an output vector,

[0026] in the learning model, the combination function being capable of calculating an average of the output vectors from each first neural network and capable of providing said average as input to the second neural network; and - the step of determining the class associated with the image comprises the following sub-steps: • providing the image to be classified as input to the classification model to calculate a vector associated with the image to be classified, comprising a plurality of components corresponding respectively to a class from the set of predefined classes, • for each component of the calculated vector, comparison of said component with a predefined threshold, • if one of the components is greater than the predefined threshold, then determination (158) of the class associated with the image to be classified as being the class corresponding to said component, and if none of the components is greater than the predefined threshold, then determination that the image to be classified does not belong to any class in the set of predefined classes.

[0027] The present invention also relates to a set of computer program product(s) comprising software instructions which, when executed by one or more computers, implement a classification method as defined above.

[0028] The present invention also relates to an electronic system for classifying an image to be classified among a set of predefined classes, from a classification model, the system comprising an electronic learning device comprising: a module for training a learning model from a first neural network and a second neural network, an input module of a training dataset of the learning model, a module for applying a machine learning algorithm to the learning model, from the learning data, to obtain a first neural network trained from the first neural network, and a second neural network trained from the second neural network, a module for generating the classification model by connecting trained neural networks,

[0029] the classification system further comprising an electronic operating device comprising: a module for receiving the image to be classified, a module for determining the class associated with the image to be classified by providing the image to be classified as input to the classification model,

[0030] in which the training module is capable of training the learning model by connecting a predefined combination function connected to the output of each of the two first neural networks, and by connecting the second neural network to the output of the predefined combination function, each training data item comprising a pair of distinct images each for a first neural network,

[0031] the application module being specific to the fact that throughout the application of the automatic learning algorithm, the first neural networks of the learning model are identical to each other, and

[0032] the generation module is capable of generating the classification module.

[0033] These characteristics and advantages of the invention will appear on reading the description which follows, given by way of non-limiting example, and made with reference to the appended drawings, in which: - [Fig.l] [Fig.l] is a schematic view of an electronic classification system according to the invention; - [Fig.2] [Fig.2] is a diagram representing a learning model formed by the electronic system of [Fig.l]; - [Fig.3] [Fig.3] is a diagram representing a classification model generated by the electronic system of [Fig.l]; - [Fig.4] [Fig.4] is a representation of a latent space resulting from a first neural network to which training is applied by the electronic system of [Fig.l]; and - [Fig.5] [Fig.5] is a flowchart of a classification method according to the invention, implemented by the classification system of [Fig.l].

[0034] [Fig.l] represents an electronic system 10 for classifying an image to be classified 12. The classification system 10 comprises an electronic learning device 15 and an electronic operating device 20 connected to each other.

[0035] Each of the devices 15, 20 is preferably a computer. For this purpose, the learning device 15 comprises, for example, a first processor 25 and a first memory 30 comprising software instructions suitable for being executed by the first processor 25. Similarly, the operating device 20 comprises a second processor 35 and a second memory 40 also comprising software instructions suitable for being executed by the second processor 35.

[0036] As an optional addition, the operating device 20 preferably comprises a display screen 41, also called a monitor, capable of displaying content intended for a user of the operating device 20.

[0037] The first memory 30 of the learning device 15 comprises an input module 45, a module 50 for training a learning model, a module 55 for applying a machine learning algorithm to the learning model, and a module 60 for generating a classification model.

[0038] For its part, the second memory 40 comprises a module 65 for receiving the image to be classified 12, a module 70 for determining the class associated with the image to be classified 12, and optionally a display module 71 on the monitor 4L

[0039] In the example of [Fig.l], the modules 45, 50, 55, 60, 65, 70, 71 are each produced in the form of one or more software programs, or a software brick, executable by the respective processor 25, 35.

[0040] Each software is further capable of being recorded on a medium, not shown, readable by computer.

[0041] The computer-readable medium is, for example, a medium capable of storing electronic instructions and of being coupled to a bus of a computer system. For example, the readable medium is an optical disk, a magneto-optical disk, a ROM memory, a RAM memory, any type of non-volatile memory (for example EPROM, EEPROM, FLASH, NVRAM), a magnetic card or an optical card. On the or each readable medium is then stored a computer program comprising software instructions. It is then understood that we are talking about a set of computer program products comprising a learning computer program product, corresponding to the electronic learning device 15, and an operating computer program product corresponding to the electronic operating device 20.

[0042] In a variant not shown, each device 15, 20, or even each module 45, 50, 55, 60, 65, 70, 71, is each produced in the form of a programmable logic component, such as an FPGA (Field Programmable Gate Array) or even an integrated circuit such as an ASIC (Application Specific Integrated Circuit).

[0043] The input module 45 is capable of receiving a first neural network 75 and a second neural network 80.

[0044] In a manner known per se, the first 75 and the second 80 neural networks are artificial neural networks respectively comprising an input layer comprising a plurality of neurons. Each neural network 75, 80 further comprises one or more hidden layers respectively comprising a plurality of neurons, and an output layer also comprising a plurality of neurons. Each neural network 75, 80 further comprises connections between the neurons of the different respective layers, each connection having an adjustable synaptic weight, also called adjustable weight.

[0045] The input module 45 is suitable for receiving the first 75 and second 80 neural networks in a naive state. The expression “a neural network in a naive state” means a neural network that has not yet undergone a learning step. In other words, each adjustable synaptic weight of each neural network 75, 80 is set to an arbitrary value, for example equal to 1. In particular, in [Fig. 1], the naive type of the neural networks 75, 80 is represented by the fact that the connections are in dotted lines and not in solid lines.

[0046] Preferably, the first neural network 75 is an encoder. The first neural network is configured to, once training is applied to it, convert a digitized image into a vector, called the output vector 82. The output vector 82 is representative of the content of the image. In other words, when an image is provided as input to the first neural network 75 having undergone training, the first neural network 75 is capable of providing an output vector 82 representative of the elements illustrated in the image.

[0047] It is also said that the output vector 82 is a representation of the image provided as input to the first neural network 75, in a latent space. In the latent space, two algebraically close vectors are representative of substantially similar image contents.

[0048] The second neural network 80 is preferably a classifier. The second neural network 80 is configured to, once learning is applied to it, determine, from an output vector 82, a class to which the content of the image associated with the output vector 82 belongs, from among a set of predefined classes.

[0049] For example, the set of predefined classes includes the following classes: a drone, an airplane, and a helicopter. It is then understood that the classifier 80 is capable of identifying, from the output vector 82 associated with an image, whether said image represents a drone, an airplane, or a helicopter, or none of them.

[0050] The input module 45 is further capable of obtaining a set of unrepresented learning data.

[0051] For this purpose, the input module 45 is capable of receiving, from an external device and / or user, a set of labeled data 85. Each labeled data item 85 comprises an image 90 and a vector 95.

[0052] In the example of predefined classes explained above, the labeled data set 85 includes images 90 representing a drone, images 90 representing an airplane, images 90 representing a helicopter, and images 90 representing something else.

[0053] Each vector 95 comprises a plurality of components. Each component of the vectors 95 is associated with a predefined class. In the example described above, the vector 95 then comprises three components, i.e. one component for each: a drone, an airplane, and a helicopter.

[0054] For each labeled data item 85, the vector 95 associated with the image 90 comprises a first value for the component corresponding to the class represented in the image 90 and a second value for all the other components. The first value is for example equal to 1, the second value being for example equal to 0. In other words, if the image 90 represents a drone, the associated vector 95 comprises the value 1 for the component associated with the drone class and the value 0 for each other component.

[0055] The input module 45 is suitable for forming, from the labeled data set 85, the learning data set. For this, the input module 45 is suitable for forming pairs of images 90, and associating with each pair of images 90 a class vector comprising a number of components equal to the number of classes. Each component of each class vector is equal to the first value if each of the two images 90 forming the training data represents an object corresponding to the class associated with said component, or the second value otherwise. In other words, if, for any training data, the two images 90 represent a drone, then the class vector will have a component associated with the “drone” class having the first value, the other components having the second value.

[0056] According to one embodiment, when the image 90 does not correspond to any of the predefined classes, each coefficient of the associated vector 95 has a third value. The third value is for example equal to the quotient between the value 1 and the number of classes. In other words, if the number of classes is equal to three, the third value is equal to 1 / 3.

[0057] Alternatively, the third value is equal to the second value.

[0058] As previously, if the two images 90 of any learning data do not represent an object belonging to the same class, then all the components of the class vector are identical to each other, for example equal to the third value.

[0059] The training module 50 is capable of training a learning model 100, represented in [Fig.2], from the first 75 and second 80 neural networks.

[0060] As illustrated in [Fig.2], the learning model 100 comprises twice the same first neural network 75, a predefined combination function 105 connected to the output of the first neural networks 75, and the second neural network 80 connected to the output of the predefined combination function 105.

[0061] For the sake of clarity, the first two neural networks of the learning model 100 are denoted 75_1 and 75_2 in the remainder of the description.

[0062] According to the exemplary embodiment, the predefined combination function 105 is capable of calculating the average of the components of two vectors. In other words, the combination function 105 is capable of summing two vectors to form a resulting vector and capable of dividing each component of the resulting vector by two. In other words, the predefined combination function 105 complies with the following equation:

[0063] [Math.l] / («,

[0064] where a and b are any vectors of the same dimensions, and f is the predefined combination function.

[0065] In other words, the predefined combination function 105 performs an average of the coefficients of the output vectors 82 from the respective first neural networks 75_1, 75_2.

[0066] The application module 55 is capable of applying a machine learning algorithm to the learning model 100, from the learning data obtained by the input module 45. The machine learning algorithm is for example a supervised machine learning algorithm. In other words, the machine learning algorithm aims to train the learning model 100 so that, when a pair of images 90 of the learning data is provided to it as input, via the first neural networks 75_1, 75_2, said learning model 100 provides as output of the second neural network 80, a vector substantially similar to the class vector of said learning data.It is clear that providing the images 90 of the training data to the first neural networks 75_1, 75_2 means that one of the images 90 is provided to one of the first neural networks 75_1 and the other image 90 is provided to the other first neural network 75_2. By "training the learning model 100" is meant the adjustment of the adjustable synaptic weights present on the connections between the neurons of the first 75_1, 75_2 and between the neurons of the second 80 neural networks.

[0067] In a manner known per se, the machine learning algorithm is an optimization algorithm aimed at minimizing the differences between the class vectors and the vectors resulting from the second neural network 80, for all the pairs of images 90 of the training data, by modifying the adjustable synaptic weights of the first 75_1, 75_2 and second 80 neural networks. For example, the optimization algorithm is a stochastic gradient descent algorithm known per se.

[0068] The application module 55 is specific so that, throughout the application of the automatic learning algorithm, the first 75_1, 75_2 neural networks are identical to each other. In other words, the optimization algorithm is only applied to the adjustable weights of one of the first neural networks 75_1 and to the adjustable weights of the second neural network 80. Thus, the adjustable weights of the other first neural network 75_2 are, at each instant, a copy of the adjustable weights of the first neural network 75_1 on which the optimization algorithm is applied, throughout the application of the learning algorithm.

[0069] The application module 55 is configured to obtain, at the end of the application of the automatic learning algorithm, a first trained neural network 75* from the first neural network 75, and a second trained neural network 80* from the second neural network 80.

[0070] Again with reference to [Fig. 1], the generation module 60 is capable of generating a classification model 110 by connecting the second trained neural network 80* to the output of the first trained neural network 75*, as illustrated in [Fig. 3],

[0071] In [Fig. 3], the classification model 110 is represented by concatenation of the first 75* and second 80* trained neural networks. The “trained” nature of the neural networks 75*, 80* is represented in [Fig. 1] by the fact that the connections between the neurons in the classification model 110 are in solid lines. The learning device 15 is configured to transmit the classification model 110 to the operating device 20. This transmission is for example carried out via a wired connection between the devices 15, 20 or by a wireless connection.

[0072] According to one embodiment, the electronic operating device 20 is on board a vehicle, such as an aircraft. Thus, the transmission of the classification model 110 corresponds to a loading of said model 110 into the device 20, for example prior to the loading of said operating device 20 on board the vehicle.

[0073] The operating device is for example connected to a camera on board the vehicle, not shown. The camera is then configured to capture a plurality of images at successive times and to transmit the captured images to the operating device. The captured images then include the image to be classified 12.

[0074] The reception module 65 is configured to receive the image to be classified 12, for example from the camera not shown.

[0075] Alternatively, the receiving module 65 is configured to receive the image to be classified 12 from the same or another external user, for example via non-represented peripherals, from the operating device 20.

[0076] The determination module 70 is capable of determining the class associated with the image to be classified 12 received by the reception module 65. For this, the determination module 70 is capable of providing the image to be classified 12 as input to the classification model 110 obtained from the learning device 15. The determination module 70 is thus configured to calculate, from the classification model 110, a respective vector 112 of the image to be classified 12. The calculated vector 112 is the vector from the second trained neural network 80* in the classification model 110.

[0077] The calculated vector 112 comprises a plurality of components corresponding respectively to the classes of the set of predefined classes. Preferably, each component of the calculated vector 112 has a value between the second predefined value and the first predefined value.

[0078] Optionally, the determination module 70 is capable of comparing each component of the calculated vector 112 to a predefined threshold. If one of the components is greater than the predefined threshold, the determination module 70 is then capable of associating with the image to be classified 12, the class corresponding to said corresponding coefficients. If none of the components of the calculated vector 112 is greater than the predefined threshold, the determination module 70 is capable of not associating, with the image to be classified 12, any class from the set of predefined classes.

[0079] The value of the predefined threshold is between the first value and the second value.

[0080] The value of the predefined threshold results from a compromise. Indeed, the value of the predefined threshold is high enough to properly discriminate images that do not represent an object belonging to any class. It is low enough to prevent images representing objects associated with one of the predefined classes from being considered as not belonging to any of the predefined classes. However, to limit the risk of misclassification, it is preferable for the value of the predefined threshold to be too high rather than too low.

[0081] For example, if all the images representing objects of one of the classes are classified with a probability greater than 0.70 and all the images representing objects of none of the classes are classified with a probability less than 0.40, then the threshold value is located between the two values. Preferably, in this case, the threshold value is set at 0.55, i.e. a value less than 0.7 and greater than 0.4.

[0082] The optional display module 71 is capable of sending to the monitor 41 an instruction to display a message indicating the determined class of the image to be classified 12. In other words, the display module 71 is capable of, if a class from the set of predefined classes has been determined by the determination module 70, sending to the monitor 41 the display instruction indicating said class. In the above-mentioned example, the display instruction aims to display a message of the type: “the image to be classified represents a drone”, “the image to be classified represents a helicopter”, or “the image to be classified represents an airplane”.

[0083] The display module 71 is capable of, if the determination module 70 has not determined any class associated with the image to be classified 12, sending to the monitor 41 the instruction to display a message of the type: “the image to be classified 12 does not correspond to any of the predefined classes”.

[0084] [Fig.4] is a schematic representation of the latent space at the output of the first trained neural network 75*. Thus, each output vector 82 of the first trained neural network 75* is represented therein.

[0085] For the sake of understanding, the latent space is represented in [Fig. 4] in two dimensions, thus presupposing that each output vector 82 consists of two components. However, this presupposition is only taken for the purposes of illustration of [Fig. 4]. It is clear that the output vectors 82 have a number of components generally greater than two.

[0086] In [Fig.4], the output vectors 82_1 represented by stars correspond to output vectors from images representing an object belonging to the same first class. Similarly, the output vectors 82_2 represented by a solid circle correspond to the output vectors associated with images representing an object of a second class.

[0087] Thus, in [Fig.4], the contour 113_1 delimits a zone of the latent space for which any output vector 82_1 is inside this contour and is determined as belonging to the first class by the determination module 70. Similarly, the contour 113_2 delimits a zone of the latent space for which any output vector 82_2 being inside this contour is determined as belonging to the second class by the determination module 70.

[0088] The output vectors 82_3, 82_4, 82_5 represented by rectangles, squares, and triangles are output vectors 82_3, 82_4, 82_5 associated with images representing an object not belonging to any of the predefined classes.

[0089] In particular, the output vectors 82_3 represented in rectangular form correspond to adversarial examples. In other words, these rectangular-shaped output vectors 82_3 correspond to images that have been modified to try to deceive the classification model 110. These output vectors 82_3 are, for example, images representing content associated with the first predefined class but modified to be classified as belonging to the second predefined class by the classification model 110. It can be clearly seen that these output vectors 82_3 are not associated with either the first or the second class.

[0090] The output vectors 82_4, 82_5 represented in the form of a square or a triangle correspond to OOD images. In other words, these images do not represent an object belonging to one of the predefined classes or substantially similar to objects belonging to these predefined classes. For example, if the first class corresponds to drones and the second class corresponds to airplanes, said output vectors 82_4, 82_5 correspond to images representing birds, helicopters or other objects.

[0091] It is clear that the learning model 100 comprising two first neural networks 75_1, 75_2 makes it possible to carry out learning of said first neural networks 75_1, 75_2 and of the second neural network 80, processing in a similar manner the case where one of the two images 90 of the learning data does not belong to any of the predefined classes and the case where the two images 90 of the learning data do not represent objects of the same class.

[0092] As indicated previously, the output vector 82 from each of the first two neural networks 75_1, 75_2 following learning, is representative of the object present on the respective image supplied to said first neural networks 75_1, 75_2. Thus, when two distinct images representing objects belonging to the same class, are provided as input to the first two neural networks 75_1, 75_2, the output vectors from the first neural networks 75_1, 75_2 are substantially similar. The combination function 105 averaging the two output vectors 82, the vector from the combination function 105 is also substantially similar to each of the two output vectors 82 provided to it as input. It is then easy for the classifier 80 to classify the vector obtained as output from the combination function 105 as belonging to the predefined class associated with the images provided as input to the learning model 100.

[0093] Conversely, when the two images 90 provided as input to the first two neural networks 75_1, 75_2 following learning represent objects belonging to two distinct predefined classes or to none of them, the associated output vectors 82 are not substantially similar. Thus, the application of the combination function 105 to these two output vectors 82 provides a vector which is not comparable to any of said output vectors 82. The application of the second neural network 80 to this vector resulting from the combination function 105 then makes it easy to classify this vector as associated with none of the predefined classes.

[0094] In [Fig.4], this phenomenon is characterized by a substantially wide band between the first contour 113_1, and the second contour 113_2.

[0095] The operation of the classification system 10 will now be described with reference to [Fig.5] representing a flowchart of a method 115 for classifying the image to be classified 12.

[0096] The classification method 115 comprises a learning phase 120 then an exploitation phase 122.

[0097] The learning phase 120 comprises a step 125 of training the learning model 100. During this training, the input module 45 receives the first 75 and the second 80 neural networks. Then, the training module 50 trains the learning model 100 by connecting the predefined combination function 105 to the output of twice the first neural network 75_1, 75_2, and by connecting the second neural network 80 to the output of the predefined combination function 105. During the training step 125, each of the first 75_1, 75_2 and second 80 neural networks is naive.

[0098] Then, the learning phase 120 comprises a step 130 of obtaining the learning data set. As explained previously, during the obtaining step 130, the input module 45 receives the labeled data 85 and forms the learning data set from said labeled data 85.

[0099] The learning phase 120 further comprises a step 135 of applying the automatic learning algorithm to the learning model 100. For this, the application module 55 adjusts the adjustable synaptic weights of the first 75_1, 75_2 and second 80 neural networks such that, for each training data item, the supply of each of the two images 90 to a respective first neural network 75_1, 75_2 makes it possible to obtain, at the output of the training model 100, a vector substantially identical to the class vector of the training data item. Thus, at the end of the application step 135, the application module 55 then obtains the first trained neural network 75* and the second trained neural network 80*.

[0100] The learning phase 120 then comprises a step 140 of generating the classification model 110. During this step 140, the generation module 60 connects the second trained neural network 80* to the output of the first trained neural network 75*.

[0101] Optionally, the learning phase 120 further comprises a transmission step 145 during which the generation module 60 transmits the generated classification model 110 to the electronic operating device 20. As explained previously, this transmission is for example carried out via a wireless or wired connection between the electronic learning device 15 and the electronic operating device 20.

[0102] The operating phase 122 is implemented at the end of the learning phase 120. Thus, during the implementation of the operating phase 122, the electronic operating device 20 comprises the classification model 110.

[0103] The operating phase 122 comprises a step 150 of receiving the image to be classified 12, by the on-board camera not shown for example.

[0104] The exploitation phase 122 further comprises a step 155 of determining the class associated with the image to be classified 12.

[0105] For this purpose, the determination step preferably comprises a supply sub-step 156 during which the determination module 70 supplies the image to be classified 12, received by the reception module 65, to the classification model 110, i.e. as input to the first trained neural network 75* of the classification model 110 to calculate the calculated vector 112.

[0106] Then, the determination step 155 comprises a comparison sub-step 157 during which the determination module 70 compares each component of the calculated vector 112, with the predefined threshold.

[0107] If during the comparison sub-step 157, one of the components of the calculated vector 112 is greater than the predefined threshold, then, during a first determination sub-step 158, the determination module 70 determines that the class associated with the image to be classified 12 is the class corresponding to said component greater than the predefined threshold.

[0108] Conversely, if during the comparison sub-step 157, none of the components of the calculated vector 112 is greater than the predefined threshold, then during a second sub- determination step 159, the determination module 70 determines that the image to be classified 12 is not associated with any of the predefined classes.

[0109] Preferably, the operating phase 122 further comprises a display step 160 during which the display module 71 sends, to the monitor 41 of the operating device 20, the display instruction described previously.

[0110] According to a variant, during the training step 125, the input module 45 receives the first 75 and second 80 neural networks in a pre-trained state, rather than in the naive state. In other words, prior to the reception by the input module 45 of the first 75 and second 80 neural networks, training has already been carried out on the first 75 and second 80 neural networks. Thus, the adjustable synaptic weights of the neural networks 75, 80 are not all equal to each other when they are received by the input module 45.

[0111] As explained with reference to [Fig.4], the training of the learning model 100 comprising twice the first neural network 75_1, 75_2 makes it possible to generate a classification model 110 capable of discriminating between adversarial examples and OODs. Thus, the classification model 110 is more robust while remaining embeddable in a mobile system.

Claims

1. Claims Method (115) for classifying an image to be classified (12) among a set of predefined classes, from a classification model (110), the method (115) being implemented by an electronic classification system (10), the method (115) comprising a learning phase (120) comprising the following steps: - training (125) a learning model (100) from a first neural network (75) and a second neural network (80), - obtaining (130) a training data set of the training model (100), - applying (135) a machine learning algorithm to the learning model (100), from the learning data, to obtain a first trained neural network (75*) from the first neural network (75), and a second trained neural network (80*) from the second neural network (80), - generation (140) of the classification model (110) by connecting the trained neural networks (75*, 80*), the method (115) further comprising an operating phase (122) comprising the following steps: - reception (150) of the image to be classified (12), - determination (155) of the class associated with the image to be classified (12) by providing the image to be classified (12) as input to the classification model (110), characterized in that during the training step (125), the trained learning model (100) comprises twice the same first neural network (75_1, 75_2), a predefined combination function (105) connected to the output of the first neural networks (75_1, 75_2), and the second neural network (80) connected to the output of the predefined combination function (105), each training data comprising a pair of distinct images (90) each for a first neural network (75_A, 75_2), during the application step (135) of the machine learning algorithm, the first neural networks (75_1, 75_2) of the learning model (100) being identical to each other, and during the generation step (140), the classification model (110) comprising a single first trained neural network (75*), in which each training data item further comprises a class vector associated with the pair of images (90), during the application step (140), the learning algorithm being such that, following its application, the learning model (100) is capable of providing a vector substantially identical to the class vector when the pair of images (90) is provided as input to said learning model (100), in which the class vector comprises a plurality of components each associated with a class such that: - if the two images (90) belong to the same class,the component associated with said class has a first predefined value preferably equal to 1, and all the other components have a second predefined value preferably equal to 0, and - otherwise, all the components are identical to each other, preferably equal to the ratio between 1 and the number of classes.,

2. Method (115) according to the preceding claim, in which the image to be classified (12) represents a drone, an airplane or a helicopter, the set of classes preferably comprising at least one of the following classes: - a drone, - an airplane, and - a helicopter.

3. Method (115) according to the preceding claim, in which the first trained neural network (75*) is an encoder capable of converting an image (90, 12) into an output vector (82) representative of the content of the image (90, 12), the second trained neural network (80*) being a classifier suitable for determining, from an output vector (82) representative of the content of an image (90, 12), a class to which the content of the image (90, 12) belongs among the set of predefined classes.

4. Method (115) according to any one of the preceding claims, in which each first (75, 75_1, 75_2) and second neural network (80) comprises several layers of neurons and adjustable synaptic weights respectively connecting two neurons of two successive layers, the step of applying (140) the machine learning algorithm comprising the adjustment of the adjustable synaptic weights of the first (75) and second (80) neural networks such that, for each training data item, the supply of each image (90) to a respective first neural network (75_1, 75_2) makes it possible to obtain at the output of the training model (100), a vector substantially identical to the class vector of the training data item.

5. Method (115) according to any one of the preceding claims, in which each first neural network (75_1, 75_1) is capable of providing, at its output, an output vector (82), in the learning model (100), the combination function (105) being capable of calculating an average of the output vectors (82) from each first neural network (75_1, 75_2) and capable of providing said average as input to the second neural network (80).

6. Method (115) according to any one of the preceding claims, in which the step of determining (155) the class associated with the image comprises the following sub-steps: - providing (156) the image to be classified (12) as input to the classification model (110) to calculate a vector (112) associated with the image to be classified (12), comprising a plurality of components corresponding respectively to a class from the set of predefined classes, - for each component of the calculated vector (112), comparison (157) of said component with a predefined threshold, • if one of the components is greater than the predefined threshold, then determination (158) of the class associated with the image to be classified (12) as being the class corresponding to said component, and • if none of the components is greater than the threshold predefined, then determination (159) that the image to be classified (12) does not belong to any class of the set of predefined classes.

7. A set of computer program product(s) comprising software instructions which, when executed by one or more computers, implement a classification method (115) according to any one of the preceding claims.

8. System (10) for classifying an image to be classified (12) among a set of predefined classes, from a classification model (110), the system comprising an electronic learning device (15) comprising: - a training module (50) for a learning model (100) from a first neural network (75) and a second neural network (80), - an input module (45) for a set of training data for the learning model (100), - a module (55) for applying a machine learning algorithm to the learning model (100), from the training data, to obtain a first trained neural network (75*) from the first neural network (75), and a second trained neural network (80*) from the second neural network (80), - a generation module (60) for the classification model (110) by connection trained neural networks (75*, 80*),the classification system (10) further comprising an electronic operating device (20) comprising: - a module (65) for receiving the image to be classified (12), - a module (70) for determining the class associated with the image to be classified (12) by providing the image to be classified (12) as input to the classification model (110), characterized in that the training module (50) is capable of training the learning model (100) by connecting a com- function, predefined combination (105) connected at the output of each of twice the first neural network (75_1, 75_2), and by connecting the second neural network (80) at the output of the predefined combination function (105), each training data comprising a pair of distinct images (90) each for a first neural network (75_A, 75_2), the application module (55) being suitable for ensuring that throughout the application of the automatic learning algorithm, the first neural networks (75_1, 75_2) of the learning model (100) are identical to each other, and the generation module (60) is suitable for generating the classification module (110) in which each learning data item further comprises a class vector associated with the pair of images (90), the learning algorithm being such that, following its application by the application module (55), the learning model (100) is capable of providing a vector substantially identical to the class vector when the pair of images (90) is provided as input to said learning model (100), wherein the class vector comprises a plurality of components each associated with a class such that: - if the two images (90) belong to the same class, the component associated with said class has a first predefined value preferably equal to 1, and all the other components have a second predefined value preferably equal to 0, and - otherwise, all components are identical to each other, preferably equal to the ratio between 1 and the number of classes