Method and device for determining at least one representation of an image

Abstract

Method and device for determining at least one representation of an image. This disclosure relates to a method for determining at least one representation of a target image of a target domain from an image representation of at least one source domain in a mobile system, said representation being intended for the identification of objects in a target image of a scene, captured by an image acquisition device collaborating with the mobile system, the method comprising: - encoding a plurality of text pairs, each pair comprising a text of said source domain and a text of said target domain, by a neural network-based encoder comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, said image encoder comprising a plurality of sequential functions separated into at least two groups, configured to encode an image,- Obtaining a representation of the image in a space other than the latent space. Figure for the abbreviation: Fig. 5.

Description

Title of the invention: Method and device for determining at least one representation of an image. Technical field

[0001] This disclosure relates to the fields of image processing and analysis, as well as computer vision. More specifically, this disclosure concerns a method and device for configuring a scene understanding neural network applied to real data, when only synthetic data is available for training. This disclosure finds a particularly advantageous, though not limiting, application in the implementation of embedded systems, such as autonomous vehicles or navigation systems. Previous technique

[0002] In many fields, real data is not used when training the algorithms that will subsequently analyze it. This is because it is either unavailable or expensive to obtain. Algorithms are therefore developed, or trained for neural network-based algorithms, using synthetic data. Discrepancies are often observed between real and synthetic data. This discrepancy is problematic because an algorithm trained on synthetic data may see its performance decrease once deployed on real data. This problem applies to any discrepancy between training and test data.

[0003] Such algorithms can be used for scene perception in autonomous vehicles or for driver assistance or decision-making systems in a mobile system. Indeed, the operation of autonomous vehicles or driver assistance or decision-making systems in a mobile system requires good perception of the environment in which the vehicle is operating. This involves detecting obstacles in the environment, analyzing the vehicle's environment, and, depending on the type of vehicle, detecting pedestrians, vehicles, buildings, and traffic lights. For example, this might involve detecting objects on the ground or assessing the hazards of a landing site for an aircraft. It also requires classifying the detected objects and segmenting them semantically.Of course, databases of real and representative images of every type of environment that the mobile system might subsequently perceive cannot be exhaustive. Therefore, synthetic databases, better representative of the different situations that it will encounter, are necessary. Mobile systems must be used to train these algorithms before being put on the market.

[0004] However, even though powerful tools exist for synthesizing images as faithfully as possible to reality, qualitative differences remain between synthetic and real images. Such differences may, for example, be due to inaccuracies in texture or lighting in synthetic images. Algorithms trained on synthetic images may therefore be less effective when deployed on real images.

[0005] There is therefore a need to improve the performance of algorithms trained on synthetic data. Description of the invention

[0006] This disclosure relates to a method for determining at least one representation of a target image of a target domain from a representation of images of at least one source domain in a mobile system, said representation being intended for the identification of objects in a target image of a scene, captured by an image acquisition device collaborating with the mobile system, the method comprising: - an encoding of a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a neural network-based encoder comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, said image encoder comprising a plurality of sequential functions separated into at least two groups, configured to encode an image, - training of a first neural network (F) to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance between said texts from the target domain and the source domain of the same pair encoded by said text encoder, - training a second neural network (G) to determine, for an image of the source domain, parameters of a second augmentation function minimizing a second distance between - the encoding by the second group of sequential functions of said image encoded by the first group of sequential functions and transformed by the second augmentation function and - the transformation by the first augmentation function of the encoding of said target image by the result of the sequential encoding of said target image by the first group of sequential functions and then by the second group of sequential functions, - the determination of at least one representation of the image of the target domain being obtained as the application of said second augmentation function, resulting from said training, to the image of the target domain encoded by the first group of sequential functions.

[0007] According to some embodiments, the process includes determining a second representation, said second representation being obtained as the application of said augmentation function, resulting from said training, to the image of the target domain encoded by the first group of sequential functions and then by the second group of sequential functions.

[0008] According to certain embodiments, the process includes the determination of n representations, the representation of rank n being obtained as the application of said augmentation function, resulting from said training, to the image of the target domain encoded by the first group of sequential functions and then by the following n-1 groups of sequential functions, when said said image encoder comprises a plurality of sequential functions separated according to at least n groups.

[0009] According to certain embodiments, the first distance D is expressed according to the following formula: D=

[0010] In which fg represents the first augmentation function And represents the text encoder. Ps represents the source text. Pc represents the target text.

[0011] According to certain embodiments, said first distance D is expressed according to the following formula: D = \\UET(ps) )-Et(pc) ||2 + a( IlEt(pj ||2- 114,(^) ) ||2)2

[0012] In which fe represents the first augmentation function And represents the text encoder Ps represents the text of the source domain, Pc represents the text of the target domain, 2 represents a positive scalar

[0013] According to some embodiments, said first augmentation function is initialized as a zero linear transformation.

[0014] According to certain embodiments, said second distance is expressed by the following formula:

[0015] minllE^fEÎXx') ) )||

[0016] In which, for an image x\ g^ represents the second augmentation function, for said image x', represents the second group of sequential functions, E^ represents the first group of sequential functions, f 0 represents the first augmentation function

[0017] According to certain embodiments, when images of n source domains are available, n being greater than or equal to two, -the said training of a first neural network (F) to determine parameters of a first augmentation function by respective domain, transforms the texts of the source domain into texts of the target domain, this training being done by minimizing a first distance D between the text of the target domain encoded ET ( pc ) and the text of the source domain of the same pair, encoded by the text encoder (ET) and augmented by the first augmentation function f (ET( ps)),

[0018] According to some embodiments, for an image of a determined source domain, the first augmentation function / ^E^p^ ) is selected, whose parameters have been determined for said determined source domain, - the transformation by the first augmentation function of the encoding of said image of the source domain by the result of the sequential encoding of said image of the source domain by the first group of sequential functions and then by the second group of sequential functions, using the first selected augmentation function.

[0019] According to certain embodiments, - said target domain includes real images obtained by a capture device, - said source domain comprises synthetic images chosen from one or more of the following: - images obtained through simulation, - images obtained by generative artificial intelligence processes, - real images on which special effects have been applied, - images drawn by a graphic designer or a painter.

[0020] This disclosure also relates to a computer program comprising instructions for performing the steps of the process of determining at least one representation of a target image of a target domain from an image representation of at least one source domain according to any of its embodiments when said program is executed by a computer.

[0021] This disclosure also relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for performing the steps of the process of determining at least one representation of a target image of a target domain from an image representation of at least one source domain according to any of its embodiments.

[0022] This disclosure also relates to a method for semantic segmentation of at least one representation of an image from an image acquisition device of a mobile system, in which said representation is obtained by a method according to one of the embodiments of this disclosure.

[0023] This disclosure also relates to a computer program comprising instructions for performing the steps of the semantic segmentation process of at least one representation of an image from an image acquisition device of a mobile system, according to any of its embodiments when said program is executed by a computer.

[0024] This disclosure also relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for executing the steps of the semantic segmentation process of at least one representation of an image from an image acquisition device of a mobile system according to any of its embodiments.

[0025] This disclosure also relates to a mobile system for determining a representation of an image of a target domain captured by an image capture device, from an image representation of at least one source domain, said representation being intended for the identification of objects in said image of a scene captured by said capture device, said system comprising one or more processors configured together or separately for: - encode a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a neural network-based encoder comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, said image encoder comprising a plurality of sequential functions separated into at least two groups, configured to encode an image, - train a first neural network (F) to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance between said texts from the target domain and the source domain of the same pair encoded by said text encoder, - train a second neural network (G) to determine, for a source image, parameters of a second augmentation function minimizing a second distance between - the encoding by the second group of sequential functions of an image encoded by the first group of sequential functions and transformed by the second augmentation function and - the transformation by the first augmentation function of the encoding of said source image by the result of the sequential encoding of said image by the first group of sequential functions and then by the second group of sequential functions, - determine the representation of the image of the target domain being obtained as the application of said second augmentation function, resulting from said training, to the image of the target domain encoded by the first group of sequential functions.

[0026] This disclosure also relates to the use of the determination of at least one representation of an image of a target domain, according to one of the embodiments of this disclosure where the target domain represents a terrestrial environment in order to classify objects in images of the terrestrial environment, or to perform semantic detection or segmentation of said objects.

[0027] This disclosure also relates to a method for semantic segmentation of a representation of a terrestrial image from an image acquisition device of a mobile system, in which said representation is obtained by a method according to one of the embodiments of this disclosure.

[0028] This disclosure also relates to the use of the determination of at least one representation of an image of a target domain, according to one of the embodiments of this disclosure where the target domain includes mechanical or electronic parts, in order to classify wear defects or anomalies in these images of the target domain or to perform semantic detection or segmentation.

[0029] This disclosure also relates to a method for semantic segmentation of a representation of an image representative of mechanical or electronic parts from an image acquisition device of a mobile system, in which said representation is obtained by a method according to one of the embodiments of this disclosure.

[0030] This disclosure also relates to a method for determining the parameters of an image classification system for images of a target domain that have been captured by an image capture device, from an image classification of at least one source domain comprising: - an encoding of a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a network-based encoder neurons comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, -training of a first neural network to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance D between the encoded target domain text and the source domain text of the same pair, encoded by the text encoder and augmented by the first augmentation function. - a training of said classification system to determine its parameters, said training operating at least on said source domain images having been encoded by said trained image encoder and then transformed by said first augmentation function.

[0031] According to certain embodiments, said training of said classification system to determine its parameters, operates on said source domain images having been encoded by said trained image encoder and then transformed by said first augmentation function and on source domain images having been encoded by said trained image encoder.

[0032] According to certain embodiments, said source domain images having been encoded by said trained image encoder and then transformed by said first augmentation function and source domain images having been encoded by said trained image encoder, are provided iteratively, alternately or randomly, during said training of said classification system.

[0033] According to some embodiments, said first augmentation function is initialized as a zero linear transformation, that is to say the identity.

[0034] According to certain embodiments, the method includes an initialization of the classification system carried out using one of the following methods: a) a random method, b) a random method combined with pre-training from available generic data, encoded by the image encoder but not transformed by the first augmentation function, c) a method according to a) or b) followed by pre-training on source images encoded by the image encoder but not transformed by the first augmentation function, d) a method initializing a linear classifier with the weights of the "zero shot" classifier of the text-image models.

[0035] According to some embodiments, the method comprises a validation of the classification system from a second set of images from the source domain, a first set of images from the source domain being used during training. said classification system, said validation using as input the images from the second set of images from the source domain encoded by the image encoder.

[0036] According to certain embodiments, - said target domain includes real images obtained by a capture device, - said source domain comprises synthetic images chosen from one or more of the following: - images obtained through simulation, - images obtained through generative artificial intelligence processes, - real images onto which special effects have been applied - images drawn by a graphic designer or a painter.

[0037] According to certain embodiments, when images of n source domains are available, n being greater than or equal to two, - said training of a first neural network to determine parameters of a first augmentation function by respective source domain, transforms the texts of the source domain into texts of the target domain, this training being done by minimizing a first distance D between the encoded target domain text and the source domain text of the same pair, encoded by the text encoder and augmented by the first augmentation function' - said training of said classification system to determine its parameters, operates on said images of the n source domains having been encoded by said trained image encoder and then transformed by said first augmentation function corresponding to their respective domain.

[0038] This disclosure also relates to a computer program comprising instructions for performing the steps of the process of determining the parameters of an image classification system of a target domain having been captured by an image capture device from an image classification of at least one source domain according to any of its embodiments when said program is executed by a computer.

[0039] This disclosure also relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for performing the steps of the process of determining the parameters of an image classification system of a target domain having been captured by an image capture device from an image classification of at least one source domain according to any of its embodiments.

[0040] This disclosure also relates to a device for determining the parameters of an image classification system for images of a target domain that have been captured by an image capture device, based on a classification of images from at least one source domain, the device comprising one or more processors configured together or separately for instructions to execute the steps of the process according to this disclosure, according to any of its embodiments. Thus, this disclosure also relates to a device for determining the parameters of an image classification system from images of a target domain that have been captured by an image capture device, from an image classification of at least one source domain comprising one or more processors configured together or separately for - encode a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a neural network-based encoder comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, -train a first neural network to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance D between the encoded target domain text and the source domain text of the same pair, encoded by the text encoder and augmented by the first augmentation function, -train said classification system to determine its parameters, said training operating at least on said source domain images having been encoded by said trained image encoder and then transformed by said first augmentation function.

[0041] This disclosure also relates to a method for determining the parameters of a classification system for representative images of a terrestrial scene, in order to classify elements of said scene, said elements being representative of moving or stationary objects, based on an image classification of at least one source domain, the method comprising: - an encoding of a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a neural network-based encoder comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, - training of a first neural network to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance D between the encoded target domain text and the source domain text of the same pair, encoded by the text encoder and augmented by the first augmentation function, - training of said classification system to determine its parameters, said training operating at least on said images representing scenes terrestrial, having been encoded by said driven image encoder and then transformed by said first augmentation function.

[0042] This disclosure also relates to a method for determining the parameters of a classification system for representative images of mechanical or electronic parts, in order to classify wear defects or anomalies on the parts, based on an image classification of at least one source domain comprising representative images of anomalies or wear defects on said parts, the method comprising: - an encoding of a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a neural network-based encoder comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, - training of a first neural network to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance between the encoded target domain text and the source domain text of the same pair, encoded by the text encoder and augmented by the first augmentation function, - a training of said classification system to determine its parameters, said training operating at least on said representative images of mechanical or electronic parts, having been encoded by said trained image encoder and then transformed by said first augmentation function.

[0043] This disclosure also relates to the use of the method for determining the parameters of a classification system for the classification of wear defects or anomalies on mechanical or electronic parts, the source images being representative of the actual parts, after manufacture and / or after use.

[0044] This disclosure also relates to a method for classifying an image of a target domain, said images being images captured by an image capture device, said method comprising - the classification of said captured images by at least one classification system driven by a process for determining the parameters of an image classification system of a target domain having been captured by an image capture device, from an image classification of at least one source domain according to one of the embodiments of this disclosure, said classification associating a class with a captured image.

[0045] This disclosure also relates to a method for classifying representative images of mechanical or electronic parts, the images being captured by an image capture device, said method comprising: - the classification of said captured images by at least one classification system driven by a process for determining the parameters of an image classification system of a target domain having been captured by an image capture device, from an image classification of at least one source domain according to one of the disclosed embodiments, said classification associating a class with a captured image.

[0046] This disclosure also relates to a method for classifying representative images of terrestrial scenes, the images being captured by an image capture device, said method comprising: - the classification of said images captured by at least one classification system driven by a process for determining the parameters of a classification system for images captured by an image capture device and representative of a terrestrial scene, in order to classify elements of said scene, said elements being representative of moving objects, from an image classification of at least one source domain according to one of the disclosed embodiments, said classification associating a class with a captured image.

[0047] This disclosure also relates to a computer program comprising instructions for carrying out the steps of the classification process according to any of its embodiments when said program is executed by a computer.

[0048] This disclosure also relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for carrying out the steps of the classification process according to any of its embodiments.

[0049] This disclosure also relates to a device for classifying an image of a target domain, said images being images captured by an image capture device, the device comprising one or more processors configured together or separately to implement the instructions for executing the steps of the process according to this disclosure, according to any of its embodiments. Thus, this disclosure also relates to a device for classifying an image of a target domain, said target images being images captured by an image capture device, comprising one or more processors configured together or separately to - to classify said images captured by at least one classification system trained by a process for determining the parameters of an image classification system of a target domain having been captured by an image capture device, based on an image classification of at least one source domain according to one of the methods of implementing this disclosure, said classification associating a class with a captured image.

[0050] Other features and advantages of this disclosure will become apparent from the description below, with reference to the accompanying drawings which illustrate an example of an embodiment without being limiting in any way. Brief description of the drawings

[0051] [Fig-1] Figure 1 represents one embodiment of a first augmentation function,

[0052] [Fig.2] Figure [Fig.2] represents one embodiment of training a classification algorithm,

[0053] [Fig.3] Figure [Fig.3] represents one embodiment of the deployment of the classification function,

[0054] [Fig.4] Figure [Fig.4] represents an embodiment of an image encoder,

[0055] [Fig. 5] Fig. 5 represents one embodiment of obtaining a second augmentation function from a visual encoder to obtain a representation suitable for semantic detection and segmentation,

[0056] [Fig.6] The [Fig.6] represents an embodiment of a device according to an embodiment of the present disclosure.

[0057] [Fig.7] The [Fig.7] represents an embodiment of a system according to certain embodiments of this disclosure. Description of the implementation methods

[0058] This disclosure relates to machine learning algorithms, and in particular deep learning algorithms. Existing solutions assume that the distribution of training data is identical to the distribution of the data that the trained model will have to process during its deployment (inference data). In many applications, there is a discrepancy between these distributions, and, without careful attention, the performance of a learning algorithm can be negatively impacted by this discrepancy. In the remainder of this disclosure, we will refer to the training data domain as the source domain and the inference data domain as the target domain. In some cases, the difference in domain arises from the fact that there is no real data in the target domain and that the only training data comes from other domains.This disclosure aims to make an algorithm robust to a change in distribution. For example, we might want to use synthetic images to recognize real-world objects for which we don't have images.

[0059] A target domain could for example be a real environment where autonomous vehicles circulate, a set of mechanical parts to be controlled.

[0060] An associated source domain may, for example, include synthetic data representative of an environment and autonomous vehicles circulating in that environment.

[0061] According to a second example, a target domain can be a set of controlled mechanical parts in a factory under particular lighting.

[0062] An associated source domain may consist of synthetic data simulating anomalies, wear defects and other features in parts, for example under this particular lighting. The present disclosure may, for example, be applied to classification algorithms as well as to semantic detection and segmentation algorithms and in particular allows for the improvement of their robustness.

[0063] This disclosure may apply, by way of non-limiting example, to data representative of an environment in which a vehicle operates, for example.

[0064] An autonomous or intelligent land or air vehicle (in the sense that it provides assistance to a human pilot or in decision-making) needs to perceive its environment, for example, what is happening around or below it. Environmental perception includes: - detection tasks, such as detecting a pedestrian, a vehicle, or a dwelling by providing the coordinates within the image of a bounding box; - classification tasks, such as assigning a class to a bounding box like car, tank, truck, etc. - semantic segmentation tasks, such as assigning a class to each pixel in order to precisely outline elements in the image like pedestrian, car, grass, etc.

[0065] In order to perceive this environment, machine learning algorithms, for example based on neural networks, are used.

[0066] This disclosure may also apply to parts on which a conformity check must be performed.

[0067] The inspection of aeronautical parts is traditionally carried out by a human operator, either directly through visual observation of the part or by checking image recordings of the part taken by cameras or other recording devices in grayscale, color, X-ray, etc. This inspection can also be carried out automatically or semi-automatically using computer vision algorithms to classify acquisitions (normal / abnormal, type of anomaly), perform semantic detection or segmentation of abnormal areas, and classify these areas. According to this disclosure, we are working within the framework of automatic or semi-automatic inspection using statistical machine learning algorithms, for example, based on neural networks.

[0068] Machine learning algorithms intended for deployment in perception assistance systems require data for training. Data is a critical factor in the performance of a machine learning algorithm. Obtaining real-world data to train these learning algorithms can sometimes be complex. Indeed, in the case of environmental detection, it is impossible to implement segmentation algorithms in real-world conditions. Similarly, in the case of parts to be inspected, algorithms are often developed before defects occur on the actual parts. Thus, when real-world data is too expensive or complex to acquire, it becomes advantageous to generate synthetic data, which is often easier to generate in large quantities. This data is used to train the algorithms, which then operate in the real-world environment.

[0069] The following examples illustrate some applications in which the real data do not contain the information that one would nevertheless like to analyze with learning algorithms. In these cases, this information can be created in synthetic data: - When detecting objects seen from the sky by a drone, it is possible to detect all vehicles and classify them. However, some models will never have been seen in the training databases (for example, new models on the market or confidential or rare vehicles such as foreign military vehicles). - In scene segmentation for autonomous vehicles, it is difficult to predict all environments (city with skyscrapers, medium-sized city in different countries, villages, desert, etc.). Similarly, for certain use cases, an infrared camera is preferable to an RGB camera. However, infrared ranges vary, and camera qualities differ depending on the application: private vehicles, civilian, military, land, or air. Dedicated databases for these sensors are not always available and are difficult to acquire. - The factories produce a new part never produced before, so there is not enough data at the beginning of the production cycle to train a statistical learning algorithm. - Due to new regulations, a new defect must be detected if it is present. It is too costly to re-examine all the databases to see if this defect was already present previously. - We want to ensure the robustness of the algorithms to certain sensor or lighting failures, without having actually experienced such failures.

[0070] Due to the imperfect data synthesis process, in most cases, differences are observed between the synthetic and real data: style, colors, texture, etc. This discrepancy is problematic because an algorithm trained on data Synthetic models may experience a performance decrease when deployed on real-world data. This issue applies to any discrepancy between training and test data.

[0071] In the context of this disclosure, we are considering a two-dimensional data case, 2D data, which is therefore images. By real data we mean any data of a real environment acquired by a camera (RGB, infrared, etc.). By synthetic data we mean any other data: for example, data created with the help of a simulator (e.g., Blender), created using a machine learning algorithm (e.g., GAN), modified by a high-level algorithm (e.g., synthetic rain or fog overlays or overlays from other images), drawn using a graphic designer, a painter, etc.

[0072] The solution proposed in this disclosure aims to reduce the domain gap between synthetic and real data in order to minimize performance loss when a neural network is trained on synthetic data and deployed on real data. The source domain (synthetic data, available for training) and the target domain (real data, not available during training) are referred to as the source domain.

[0073] Throughout the description, source image or source domain image are used interchangeably, likewise target image or target domain image are used interchangeably.

[0074] Figure 1 represents an embodiment of learning an augmentation function fg learned to transform the representations in the latent space of a text-image neural network of texts from a source domain to a target domain.

[0075] An annotated classification database with K classes is available. This database consists of M synthetic source domains. By synthetic domain, we mean images produced by a synthetic technique. For example, a database with data produced by Blender and data produced by a GAN is considered to have two source domains. Two databases produced by the same technique but with different parameters can be considered as two domains if it is possible to describe them differently.

[0076] Textual descriptions of the source and target classes and domains are also available. For example, a textual description of a class could be its name (e.g., “pedestrian”, “car”, etc.). A textual description of a domain could be its production technique. For example, for source domains, the textual description could be: synthetic, BLENDER, graphics software, GAN, Synthetic infrared. For target domains, the following textual descriptions are possible: real, photo, realistic, infrared photo.

[0077] A text-image encoder is also available, based on a text-image neural network trained by minimizing the cosine distance between a text and an image of the same pair. This neural network typically includes at least one text encoding module and one image encoding module. For example, a visual transformer and a text transformer pre-trained with the CLIP (Contrastive Language-Image Pre-training) software can be chosen as encoders.

[0078] A set of human-imagined textual representations of one or more source domains and the target domain is also available. These textual descriptions are, for example, in a few words, the characteristic style of the images: "synthetic," "realistic," "black and white," etc. The textual descriptions are written in the language of the text used to train the visual and textual encoders.

[0079] A single textual description per domain remains insufficient to represent the full complexity and richness of a domain. Although this disclosure can function with only human-imagined textual input descriptions, the results obtained with this disclosure are more effective with an enriched representation of the domains. Thus, this disclosure may include obtaining, or enriching, a textual description of a target and source domain.

[0080] This enrichment is advantageously carried out using a text generation algorithm (for example a natural language processing (NLP) neural network) in order to generate a large number of texts (for example 100000) in the working language.

[0081] For example, if the working language is French, texts beginning with "an image of" are generated (possibly via a broad language model), where # is a word or group of words representing the classes to be classified. For example, if the aim is to classify elements found in the street, then # could be "pedestrian", "car", "sidewalk", etc. To create texts specific to the source and target domains, the word "image" in each text is replaced by the textual description of the source and target domains.

[0082] An example of generated text is "an image of a red car in the middle of the road". If the source domain is described by "synthetic image" and the target domain is described by "photograph", then the text specific to the source domain is "a synthetic image of a red car in the middle of the road" and the text specific to the target domain is "a photograph of a red car in the middle of the road".

[0083] Advantageously, words can be automatically corrected to match the grammar of the working language (for example, "de" replaced by "d'un" or "d'une" depending on the context).

[0084] In Figure 1, the AND text encoder of the text-image neural network-based encoder receives as input pairs of source texts Ps and target texts Pc. This text encoder has been previously trained; its parameters are therefore predetermined and fixed. The output of the AND text encoder is a representation Er <M des textes sources et une représentation ET ( p ) des textes cibles dans l’espace latent.

[0085] We then train a first neural network F to determine the parameters of an augmentation function f 0 transforming the texts of the source domain into texts of the target domain, this training being done by minimizing a first distance D between the text of the target domain encoded ET[ p ) and the text of the source domain of the same pair, encoded by the text encoder ET and transformed by the augmentation function fg(ET(ps)').

[0086] In one variant, F can take a predefined form (for example a linear function represented by a matrix), provided that its parameters can be learned by learning.

[0087] This distance D represents the loss function L used for training the neural network F.

[0088] According to a first embodiment, the distance D can be expressed as follows: 100891 D= \\fe(ET(py)-EApj\\;

[0090] D is preferably the L2 distance, used in regression problems. This loss function is that of the least squares problem and is convex. If fg is modeled by a linear function, then regardless of its initialization, the training of the parameters of fg converges to a unique solution.

[0091] According to a second embodiment and only if fg is modeled by a linear function, a second term D2 is added to the loss function

[0092] D2=2(||ET(P<)ll2-|| / e(Er(P5))||2)2

[0093] The distance D can be expressed aM D= ^(E^) )-ET(pc) ||\2( ||Er(p,) ||2- ||r„(£r( / >s) ) ||j2

[0094] 2 representing a positive scalar, hyperparameter of the model.

[0095] The term D2 forces f0 to remain a rotation. Since a rotation is a linear transformation that preserves norms, D2 therefore takes the form of a term of regularization penalizing the square of the difference between the standard of the source text and the standard of the augmented source text:

[0096] With the addition of the second term D2, the function fg can be initialized with the zero linear transformation, that is, the identity function. Indeed, to change a source text into a target text, it suffices to change only a small number of words. The transformation is simple, which is therefore a desirable property for the augmentation function. Furthermore, since the text-image network was trained with a cosine distance, it is preferable for the augmentation function to be close to a rotation. The zero linear transformation is both simple and a rotation, and thus constitutes a good initialization for the function fg.

[0097] Figures 2 and 3 illustrate the use of this augmentation function to transform the latent space representation of the source domain images in the text-image neural network into a target domain image representation. The transformed source image representations can then be used to train a classification algorithm.

[0098] Figure [Fig. 2] represents an example of a method for learning a classification function C, intended to classify images.

[0099] This process comprises three different functions: - the image encoder (also called the visual encoder) of the neural network-based encoder, whose parameters are fixed, - the function F dedicated to the augmentation function learned as described previously, whose parameters are fixed during this classification function learning process, - a classification function C in the training or parameter learning phase. The function C can take any form of state-of-the-art classifier (neural network, SVM, decision tree, etc.).

[0100] The image encoder E receives as input the images from one or more source domains and produces their representation in latent space. The function F associated with the source domain D receives as input the latent or vector representations E^L of the images in domain D and produces as output y'CU1 transformation by the first augmentation function fg.

[0101] The function C receives, for its training, as input, Q% of encoded and augmented target images f ( r? i* A and (100-Q)% of encoded, unaugmented target data JS)) E^Is), such that 100 > Q > 0. Augmented and unaugmented source data are iteratively provided as input for training C, either alternately or randomly. Augmented encoded source data allows C to learn to classify images from the target domain, without ever having had images from the target domain during C training. Adding a proportion strictly less than 100% of unaugmented source data during C training can allow better generalization of the classifier by mitigating possible errors in the representation of augmented source data compared to the target data.

[0102] Thus, the classification function C learns in particular to classify the encodings transformed by the image augmentation function of the source domain.

[0103] This works, firstly, because the text-image network can encode both text and images in the same latent space, and secondly, because this latent space is well-structured thanks to the use of cosine distance during its training. Thus, because the augmentation function is trained to transform text representing the source domain into text representing the target domain, it manages to transform, at least partially, an image from the source domain into its equivalent in the target domain.

[0104] Following the training of the classification function C, it is validated. This validation uses as input a second set of images from the source domain, different from the set of images from the source domain used during its training. The first set of images from the source domain can be the set of source images used for training as described above, and the second set of source images can be the set of source images used for its validation.

[0105] The second set of source images can be used for hyperparameter optimization, particularly for stopping the learning process. Furthermore, the classification function can use the E^Is representation of the second set of source images as validation data, rather than f(r(T) JA 1 s}}

[0106] This has the effect of avoiding domain overfitting and therefore being more generalist for several domains.

[0107] Indeed, the domain represented by the augmented source data remains slightly different from the target domain, so it is preferable not to overfit it.

[0108] The initialization of the classification function can be done in different ways: a) it can be carried out randomly, b) it can be performed randomly and by pre-training using available generic data (for example, from databases such as imageNet), encoded by the image encoder but not transformed by the augmentation function, c) it can be carried out according to a) or b) followed by pre-training on the source domain images encoded by the image encoder but not transformed by the augmentation function. d) Finally, it can be initialized with the so-called "zero-shot" classifier, specific to text-image models. According to this variant, the classification function is a linear function, represented by a matrix of size d * K, where d is the dimension of the latent space and K is the number of classes in the classification problem. Initialization can consist of finding the parameters of this matrix according to the following substeps: di) Create a textual description for each class in our initial classification problem in the format (if the working language is English) "a photo of a #" where # is the name of the class (e.g., car, pedestrian, etc.). d2) Encode these sentences with the text encoder of the text-image encoder to obtain a vector associated with each sentence. This gives us K vectors of dimension d. d3) Group these vectors into a matrix of size d * K. By construction, this matrix has the same dimensions as the matrix sought for the classification function.

[0109] Fig. 3 represents an example of the deployment of the classification function.

[0110] When deploying or using the classification function, the parameterized classification function as described previously. The data from the target domain Ic are encoded by the neural network-based text-image encoder and their representation E^Ic) is entered into the classification function C.

[0111] According to one embodiment, it is possible to have data from several source domains.

[0112] According to this embodiment, when data is available from several source domains, a single classification function can be determined as mentioned above from the images of the n source domains.

[0113] When images of n source domains are available, n being greater than or equal to two, -Training a first neural network to determine the parameters of a first augmentation function for each respective domain transforms the texts of the source domain into texts of the target domain. This training is done by minimizing a first distance D between the encoded text of the target domain and the text of the source domain of the same pair, encoded by the text encoder and augmented by the first augmentation function. - The training of the classification system to determine its parameters operates on the said images of the n source domains having been encoded by the image encoder trained and then transformed by the first augmentation function corresponding to their respective domain.

[0114] Figure 4 illustrates an example of the representation of an Eb image encoder.

[0115] An image encoder E! as described above can consist of a plurality of functions, and for example an alternation of linear and non-linear functions represented by parts PI, P2 and P3 in [Fig.4].

[0116] It can be noted that the visual encoder of [Fig.5] may be structurally the same as that used in Figures 1 to 3 described previously.

[0117] The encoder E comprises a plurality of linear and nonlinear functions. Such an image encoder can have different architectures: ResNet, EfficientNet, ViT (Vision Transformer)... The encoder can be separated into two parts: E^ represents the beginning of the encoder and Eb represents the remainder. In the case where the image encoder is a ResNet, i.e., a neural network consisting of four so-called "residual" blocks, E^ could, for example, be the first convolution block, and Eb is blocks 2 through 4.

[0118] The image x encoded by the encoder Ej^E^ x ) is given by the composition of E* and Eb:

[0119] E^x) = EbI(EaI(x))

[0120] The function fe as described above uses the data represented in The latent space, and not the data represented in intermediate layers such as those used, for example, by semantic detection and segmentation applications. It is therefore compatible with any task that connects to the latent space, that is, one that only needs the text-image network representations of the images.

[0121] In contrast, semantic detection and segmentation applications generally use intermediate image representations rather than the representation in latent space. For example, they may use representations 1 and 2 respectively derived from parts 1 and 2 of the encoding function.

[0122] Most semantic detection and segmentation algorithms require multiscale representations of images. To achieve this, these algorithms retrieve the outputs from different layers of the visual encoder, upstream of the neural network's latent space. Thus, the data modified by the function f0 are not used by the semantic detection and segmentation algorithms. This disclosure therefore proposes to determine a second augmentation function adapted, in particular but not exclusively, to semantic detection and segmentation. This second augmentation function can thus reduce the gap between the source and target data and improve the robustness of detection and to the semantic segmentation of target images from functions trained on source data.

[0123] Figure 5 illustrates the application of a second augmentation function to the detection and to semantic segmentation.

[0124] To modify the images upstream of the text / image joint space, a second augmentation function g^ is learned per source image x1. This function aims to mimic the action of fs on x1 by modifying the image earlier in the visual encoder. This augmentation function is inserted between E^ and and is optimized so that g^ modifies the intermediate representation of the image in order to bring it closer the representation of the image in the text-image joint space e^ of the image representation modified by f# Contrary to what has been described Previously, where one learned an augmentation function for an entire domain, here one learns a second augmentation function per image of the source domain(s).

[0125] The optimization of the second augmentation function g^ can be given by the following formula:

[0126] min||(1^(xi) ) ) - / #(£'(E?(V) ) ) ||^

[0127] A semantic detection or segmentation algorithm typically requires image representations at various locations in the network. Thus, for an image x' of the target domain, the algorithm can take as input (i) E^(x*) and (ii) Ej(E^(x') ) ■ (i) corresponds to the low-level representation and (ii) corresponds to the high-level representation of x1. With the addition of g'^,

[0128] (i) becomes g'^E" ( x' ) ) and

[0129] (ii) becomes E^g^E^x1') ) )•

[0130] Thus, both representations of x1 benefit from the increase.

[0131] According to some embodiments, the process includes determining a second representation, said second representation being obtained as the application of said augmentation function, resulting from said training, to the image of the target domain encoded by the first group of sequential functions and then by the second group of sequential functions.

[0132] According to certain embodiments, the process comprises determining n representations, the representation of rank n being obtained as the application of said augmentation function, resulting from said training, to the image of the target domain encoded by the first group of sequential functions and then by the n-1 groups of the following sequential functions, when said image encoder comprises a plurality of sequential functions separated into at least n groups.

[0133] When several source domains are available, as described above, a function fg is determined for each source domain. The function g depends on the function fg, so when a function f0 is learned for each domain, the function fg is selected from among the n functions fg.

[0134] The semantic detection and segmentation function uses (xty as input to perform semantic segmentation and detection of the image x'. If a semantic detection or segmentation method requires multiscale representations of the images, g^ is to be inserted into the image encoder in such a way that the Different representations need to be considered after the introduction of the increase in the encoder. Thus, these representations also benefit from the augmentation. For example, if the visual encoder consists of 3 blocks E1} and £^, that is: E (x) ~ EC(E>(Ea(x) ) j' Ct S' ^a^gor^^me of detection or segmentation requires as input 2 representations E^x) and E* ( Ej ( x ) ), then the function g”^ se located just after E^- Thus the detection or segmentation algorithm does indeed take into account entry only of augmented representations (i.e., involving^), which are in the case where g1^ is located just after E has . / i • g^E^x)

[0135] The processes as described above have been described using functions and are thus represented in the various Figures 1 to 5. These processes can be implemented sequentially or non-sequentially. The deployment of the classification function can be carried out following the determination of the function fg and the training of the classification function. The deployment of the semantic detection and segmentation function can be carried out as soon as the functions f 0 and g^ are determined.

[0136] Figure 6 represents a device capable of implementing the processes as described above. Such a device may have the hardware architecture of a computer and include, in particular, a processor 1, random access memory 2, read-only memory 3 and non-volatile memory 4. They also include communication means 5, for example for communicating with other devices and / or a user interface 6.

[0137] Read-only memory 3 constitutes a recording medium conforming to at least one embodiment of this disclosure, readable by the processor 1 and on which is stored a computer program PROG conforming to at least one embodiment of this disclosure, comprising instructions for The execution of process steps according to at least one embodiment of this disclosure. The PROG program defines functional modules of the device.

[0138] Fig. 7 represents an embodiment of a system implementing the embodiments described above.

[0139] It is understood from reading this disclosure that it can be implemented in single modules or devices or in several disjoint devices.

[0140] For this purpose, the method for determining the parameters can be implemented using source data as described above, and target data available for certain domains.

[0141] A 701 database may contain previously recorded source domain images, enriched according to needs and applications. The 701 database may also include source domain texts and target domain texts as described above, for one or more domains. These may be advantageously enriched. The database may also include source domain images, used particularly during the training of the classification function.

[0142] The processing device 702 may also include the text encoder ETet image Eb

[0143] Thus, the function fe can be predetermined from this target dataset, and the classification device can be trained on images from the source domains as described above. A trained classification device is then available for subsequent use, for example, in a system embedded with new target images. The device 702, whose hardware architecture can correspond to the device in [Fig. 6], is therefore configured to implement the method for determining the parameters of the target image classification system as described above, according to any one of its embodiments.

[0144] A device 704, whose hardware architecture may also take the form of a computer and correspond to the hardware architecture of [Fig. 6], may be connected to the device 702 via a network 705, which may be a cellular or any other type of Wi-Fi, LAN, or WAN network. The device 704 is connected via a wired or wireless network to an image capture device 703. In some embodiments, several image capture devices may be connected to the device 704. In some embodiments, the device 704 and the device(s) 703 form a single device.

[0145] Device 704 receives from device 702 the parameters of the classification system that it implements according to any one of the embodiments of this Disclosure. Device 702 receives target images from capture means 703 which it classifies according to this disclosure.

[0146] Device 704 can also implement the determination of the representation of one or more representations of a source image received by the capture means 703. The function can be determined by the processing device 704, from the function fg, of the received source image, when the device includes the text encoder Eret image ^.Lc device 704 can implement the semantic detection and segmentation process according to one of the embodiments described above.

[0147] This semantic classification or segmentation device 704 can be deployed in real environments as described above, for example it can be included in an embedded device, or even mobile, such as a drone, a satellite, or any object flying over the earth or another planet, or a more restricted environment, this device being able to be equipped with one or more image acquisition or capture devices 703 or being able to receive images from one or more image acquisition devices 703.

[0148] A drone can, for example, fly over an area ravaged by storms, earthquakes, fires, or other natural disasters to detect injured people and waterways that rescue services need to use. Actual data representative of the ravaged area are, of course, not available before the disaster. Synthetic data representative of the ravaged area following scenarios representative of the different disasters were used to obtain the parameters of the classification system during its training and are recorded in database 701. The classification, detection, or semantic segmentation process, as disclosed, can be carried on the flying device 704 and thus enable the classification of objects and / or their detection and semantic segmentation on the actual target images. This advantageously allows for an operational system when the synthetic data are available.Thus, it is advantageous to determine synthetic data in advance of events for seismic zones, for example, or for areas that may suffer fires, such as forests or areas particularly exposed to global warming.

[0149] Thus, this disclosure may apply to the detection of objects in a representative image of an aerial view, for example, captured using capture means onboard a flying aerial system, Device 704 and Device 703 being included in that flying aerial system. Device 702 and Database 701 may be onboard the flying or ground-based device and communicate with the flying device through Network 705.

[0150] Thus, this disclosure may relate to the use of the method for determining the parameters of a classification system for classifying terrestrial objects present in images captured by an aerial device. Terrestrial objects may be understood to mean moving or stationary objects, for example vehicles, people, buildings, animals, plants, obstacles...

[0151] To this end, the present disclosure may relate to a method for determining the parameters of a classification system for representative images of a terrestrial scene, in order to classify elements of said scene, said elements being representative of moving or stationary objects, based on an image classification from at least one source domain. The source domain may be a model or synthetic representation, or a simulation of potential terrestrial scenes comprising a synthetic, modeled, or figurative representation of the scene. The method comprises: - an encoding of a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a neural network-based encoder comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, - a training of a first neural network (F) to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance D between the text from the target domain encoded ET ( pc j and the text from the source domain of the same pair, encoded by the text encoder (ET) and augmented by the first augmentation function f (ET(ps)). - a training of said classification system to determine its parameters, said training operating at least on said representative images of terrestrial scenes, having been encoded by said trained image encoder and then transformed by said first augmentation function.

[0152] This disclosure also relates to a method for classifying representative images of terrestrial scenes, the images being captured by an image capture device, said method comprising: - the classification of said images captured by at least one classification system driven by a process for determining the parameters of a classification system for images captured by an image capture device and representative of a terrestrial scene, in order to classify elements of said scene, said elements being representative of moving objects, from an image classification of at least one source domain according to one of the disclosed embodiments, said classification associating a class with a captured image.

[0153] The classes can be representative of categories of vehicles (car, tank, truck), plants (tree, forest, flower...), people (pedestrians, child, adult), animals (dog, cat, elephant...).

[0154] This disclosure may also relate to the use of determining at least one representation of an image of a target domain, according to any embodiment of this disclosure, where the target domain represents a terrestrial environment, for the purpose of classifying objects in those images of the terrestrial environment, or for semantic detection or segmentation, for example, in the event of a natural disaster. The images may also be captured by a land vehicle, and this disclosure may be used to determine at least one representation of an image of a terrestrial environment in which a vehicle is moving, according to any embodiment of this disclosure, where the target domain represents a terrestrial environment, for the purpose of classifying objects that may constitute obstacles to the movement of the mobile vehicle in that terrestrial environment.In such an embodiment, device 704 can be in a terrestrial device, mobile or not.

[0155] This disclosure may also relate to a method for semantic segmentation of a representation of a terrestrial image from an image acquisition device of a mobile system, in which said representation is obtained by a method according to one of the embodiments of this disclosure.

[0156] In the case of an application of this disclosure to the detection of anomalies or wear defects on parts, or to the non-destructive testing of parts, the device 704 and the capture means 703 can be deployed in parts repair or maintenance facilities. The source images include synthetic images representative of anomalies or wear on the parts to be inspected.

[0157] Thus, this disclosure may relate to the use of the method for determining the parameters of a classification system for the classification of wear defects or anomalies on mechanical or electronic parts, the source images being representative of the actual parts, after manufacture and / or after use.

[0158] To this end, the present disclosure may relate to a method for determining the parameters of a classification system for representative images of mechanical or electronic parts, in order to classify wear defects or anomalies on the parts, based on an image classification of at least one source domain. The source domain may be a model or synthetic representation, or a simulation of potential anomalies or wear defects. The method comprises: - encoding a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a network-based encoder neurons comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, - a training of a first neural network (F) to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance D between the text from the target domain encoded ET(pc^ and the text from the source domain of the same pair, encoded by the text encoder (ET) and augmented by the first augmentation function). - a training of said classification system to determine its parameters, said training operating at least on said representative images of mechanical or electronic parts, having been encoded by said trained image encoder and then transformed by said first augmentation function.

[0159] This disclosure also relates to a method for classifying representative images of mechanical or electronic parts, the images being captured by an image capture device, said method comprising: - the classification of said captured images by at least one classification system driven by a process for determining the parameters of an image classification system of a target domain having been captured by an image capture device, from an image classification of at least one source domain according to one of the disclosed embodiments, said classification associating a class with a captured image.

[0160] Classes that can represent a wear defect or anomaly in the part. When a part has several defects, several classes, each corresponding to a defect, can be assigned to the same image. With such a method, it is possible to quickly detect defects on parts, provided the system has been previously trained, even when the parts are not yet in production or use.

[0161] This disclosure may also relate to the use of determining at least one representation of an image of a target domain, according to any embodiment of this disclosure, where the target domain includes mechanical or electronic parts, to classify wear defects or anomalies in those images of the target domain or to perform semantic detection or segmentation. This disclosure may be used to determine at least one representation of a representative image of mechanical or electronic parts, according to any embodiment of this disclosure, to segment or detect anomalies or wear defects on those parts.

[0162] This disclosure may also relate to a method for semantic segmentation of a representation of an image representative of mechanical or electronic parts from an image acquisition device of a mobile system, in which said representation is obtained by a method according to one of the embodiments of this disclosure.

[0163] In some embodiments, devices 702 and 704 may be a single device. This device may include the capture means 703 and the database 701. According to some embodiments, devices 701 and 702 are contained within the same device. According to some embodiments, devices 704 and 703 are contained within the same device.

Claims

1. Demands A method for determining at least one representation of a target image of a target domain from a representation of images of at least one source domain in a mobile system, said representation being intended for the identification of objects in a target image of a scene, captured by an image acquisition device collaborating with the mobile system, the method comprising: - an encoding of a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a neural network-based encoder comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, said image encoder comprising a plurality of sequential functions separated into at least two groups, configured to encode an image, - training of a first neural network (F) to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance between said texts from the target domain and the source domain of the same pair encoded by said text encoder, - training a second neural network (G) to determine, for an image in the source domain, parameters of a second augmentation function minimizing a second distance between - the encoding by the second group of sequential functions of said image encoded by the first group of sequential functions and transformed by the second augmentation function and - the transformation by the first augmentation function of the encoding of said target image by the result of the sequential encoding of said target image by the first group of sequential functions and then by the second group of sequential functions, - the determination of at least one representation of the image of the target domain being obtained as the application of said second augmentation function, resulting from said training, to the image of the target domain encoded by the first group of sequential functions.

2. A method according to claim 1 wherein said first distance D is expressed according to the following formula: \\fe(ET(p^)-ET(pc)\£ Where fy represents the first augmentation function Et represents the text encoder Ps represents the source text Pc represents the target text

3. A method according to claim 1, wherein said first distance D is expressed by the following formula: D = \\fB(ET^) )~et(,Pc) ||^4( H^) ||2- \\ffJiET(pi} ) f Where f represents the first augmentation function, Et represents the text encoder, Ps represents the text of the source domain, Pc represents the text of the target domain, and 2 represents a positive scalar

4. Method according to claim 3 wherein said first augmentation function is initialized as a zero linear transformation.

5. A method according to any one of the preceding claims in which said second distance is expressed by the following formula: ) ) - / #(Ef(£j(V) ) ) ||2 In which, for an image x', g'^ represents the second augmentation function, for said image E1^ represents the second group of sequential functions E^ represents the first group of sequential functions, fy represents the first augmentation function.

6. A method according to any one of the preceding claims, wherein, given images from n source domains, n being greater than or equal to two, said training of a first neural network (F) to determine parameters of a first augmentation function for each respective domain, transforms the texts of the source domain into texts of the target domain, this training being done by minimizing a first distance D between the text of the target domain encoded Pc) ct 'c text of the source domain of the same pair, encoded by the text encoder (ET) and augmented by the first augmentation function f ETÇp* ) ),

7. A method according to claim 6 wherein, for an image of a determined source domain, the first augmentation function f 0(ETÇp^ ) is selected whose parameters have been determined for said determined source domain, - the transformation by the first augmentation function of the encoding of said image of the source domain by the result of the sequential encoding of said image of the source domain by the first group of sequential functions and then by the second group of sequential functions, using the first selected augmentation function.

8. A method according to any one of the preceding claims wherein - said target domain comprises real images obtained by a capture device, - said source domain comprises synthetic images selected from one or more of: - images obtained by simulation, - images obtained by generative artificial intelligence processes, - real images on which special effects have been applied, - images drawn by a graphic designer or painter.

9. A method for semantic segmentation of a representation of an image from an image acquisition device of a mobile system, wherein said representation is obtained by a method according to one of the preceding claims.

10. A mobile system for determining a representation of an image of a target domain captured by an image capture device, from an image representation of at least one source domain, said representation being intended for the identification of objects in said image of a scene captured by said capture device, said system comprising one or more processors configured together or separately to: - encode a plurality of text pairs, each pair comprising a text from said source domain and a text from said target domain, by a neural network-based encoder comprising an image encoder and a text encoder, previously trained and configured to encode text and image data pairs, said image encoder comprising a plurality of sequential functions separated into at least two groups, configured to encode an image, - train a first neural network (F) to determine parameters of a first augmentation function transforming texts from the source domain into texts from the target domain, this training being done by minimizing a first distance between said texts from the target domain and the source domain of the same pair encoded by said text encoder, - train a second neural network (G) to determine, for a source image, parameters of a second augmentation function minimizing a second distance between - the encoding by the second group of sequential functions of an image encoded by the first group of sequential functions and transformed by the second augmentation function and - the transformation by the first augmentation function of the encoding of said source image by the result of the sequential encoding of said image by the first group of sequential functions and then by the second group of sequential functions, - determine the representation of the image of the target domain being obtained as the application of said second augmentation function, resulting from said training, to the image of the target domain encoded by the first group of sequential functions.

Images