Automatic recognition method and associated system
Patent Information
- Application Number
- PCT/FR2026/050151
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-02-21
- Filing Date
- 2026-02-19
- Publication Date
- 2026-08-27
Smart Images
Figure IMGF000004_0001 
Figure IMGF000010_0001 
Figure IMGF000011_0001
Abstract
Description
[0001] Description
[0002] Title: Automatic Recognition Method and Associated System
[0003] The invention relates to the field of image recognition and more particularly to the field of image recognition by machine learning techniques.
[0004] Known image recognition solutions involve several operations for which specific solutions have been developed. For example, classification aims to assign a label (or class) to an input image, while localization aims to draw a bounding box or contour delimiting one or more objects within an input image.
[0005] Detecting one or more objects in an input image generally involves these two tasks. Combining these two tasks to provide a reliable and robust solution is complex. This combination is typically performed sequentially, first by drawing a bounding box around each potential object in the input image, and then by assigning a label to each of these potential objects. Problems of this type are classified as object recognition problems.
[0006] To date, we know of learning devices capable of accurately recognizing objects using a convolutional neural network architecture. This is notably the case for two-stage detectors based on R-CNN type regions (from the English "Region-based Convolutional Neural Network").
[0007] This type of solution has the drawback of significant processing time. Improvements have been made with extensions of R-CNN called Fast R-CNN (from the English "Fast R-CNN") and then Faster R-CNN (from the English "Faster R-CNN"). However, these improved versions still involve processing times of around 200 ms per image, which remains far too slow. Furthermore, all R-CNN families have a mean average precision (inAP) of around 75% in the best-case scenario, depending on the model size (number of parameters and hyperparameters).
[0008] Other types of convolutional neural network architectures are capable of recognizing objects very quickly, such as single-stage YOLO (You Only Look Once) detectors. However, they exhibit a much lower average accuracy, on the order of 35% to 55% depending on the model size.
[0009] Processing time and accuracy therefore remain limiting factors and no solution truly provides satisfaction in this regard in the field of object recognition.
[0010] The invention improves the situation. To this end, it proposes an automatic recognition device for at least one visual signature comprising a detector arranged to receive an input image and to return, if necessary, at least one region of interest, the detector comprising a convolutional neural network arranged to receive the input image as input and to return a portion of the input image as a region of interest, the convolutional neural network having been trained to detect the potential presence of a visual signature in the image it receives as input, and an estimator arranged to receive an input region of interest and at least one image including a visual signature to be recognized, and to return a value representative of the presence of the visual signature to be recognized in the input region of interest.The estimator comprises a triple neural network configured to receive an image as input and return an associated vector as output. The triple neural network consists of three subnetworks with the same architecture and identical connection weights between layers within each subnetwork. These subnetworks include a transformer configured to receive the image and output the associated vector. The estimator is configured to determine a representation vector from the vector associated with the input region of interest and to calculate a similarity value between this representation vector and each representation vector determined from the vector associated with an image containing a visual signature to be recognized. It then returns the identifier of the image containing the visual signature for which the similarity value is highest and exceeds a chosen threshold.The device is arranged, starting from one or more input images, to call the detector to extract, if necessary, at least one region of interest for each input image, and to call the estimator with each region of interest as output from the detector, and one or more images including a visual signature to be recognized.
[0011] This device is particularly advantageous for detecting objects present in one or more images. It is both faster and more accurate than state-of-the-art solutions.
[0012] According to various embodiments, the invention may have one or more of the following characteristics:
[0013] - the estimator comprises one or more connected layers receiving the output vector from the transformer and producing a representation vector of a chosen dimension lower than that of said vector,
[0014] - The convolutional neural network is of the YOLO type or the faster R-CNN type, - The convolutional neural network is of the YOLO type only, comprising a YOLO multiplier arranged to receive the input image and return a plurality of cloned images corresponding to a combination of at least two image modification parameters, each image modification parameter being chosen from the group comprising:
[0015] * a vertical flip of the input image,
[0016] * a random translation of the input image,
[0017] * a change in the depth of the input image,
[0018] * a scaling of the input image, and
[0019] * a combination of several input images from the plurality of input images, - the convolutional neural network is of the faster R-CNN type only, comprising an R-CNN multiplier arranged to receive the input image as input and to return as output a plurality of cloned images corresponding to a combination of at least two image transformation parameters, each modification parameter being chosen from the group comprising:
[0020] * a random reframing of at least one region of interest,
[0021] * an insertion of space around the input image,
[0022] * the insertion of a space around at least one region of interest,
[0023] - The transformer used in the subnets is of the BeiT or SWIN type,
[0024] - The estimator includes a region multiplier arranged to receive as input the region of interest and to return as output a plurality of cloned regions corresponding to at least two display parameters, each display parameter being chosen from the group comprising:
[0025] * a horizontal flip of the augmented input image,
[0026] * a random resized crop of the enlarged input image, and
[0027] * a random variation of the colors of the augmented input image,
[0028] - the training phase is carried out on the internet, the training phase comprising a plurality of refinement iterations of the triple neural network, the triple neural network selecting at each refinement iteration at least one of the most relevant triplet(s) of interest from among a plurality of reference image triplets, each reference image triplet comprising the visual signature, a positive image containing the visual signature to be recognized and a negative image containing a signature distinct from the chosen visual signature, and
[0029] - Each negative image of the reference image triplet is selected according to the following optimization condition:
[0030]
[0031] < < where:
[0032] * oc is a margin of error,
[0033] * A is the visual signature to recognize.
[0034] * P is the positive image,
[0035] * N is the negative image, and
[0036] * f is a representation function.
[0037] The invention also relates to a method for the automatic recognition of at least one visual signature comprising the following operations: a) receiving one or more input images, submitting at least one or more of the input images each as input to a detector comprising a convolutional neural network arranged to return as a region of interest a portion of the input image, the convolutional neural network having been trained to detect the potential presence of a visual signature, and returning, where appropriate, at least one region of interest; b) for each region of interest of operation a) and for at least one or more images comprising a visual signature to be recognized,
[0038] bl) determine an associated vector obtained by submitting it to a triple neural network arranged to receive an image as input and to return an associated vector as output, the triple neural network comprising three subnetworks of the same architecture and having identical connection weights between layers within each respective subnetwork, which subnetworks include a transformer arranged to receive said image and to output the associated vector,
[0039] b2) determine a representation vector from the vector associated with the input region of interest and calculate a representative value of the similarity between this representation vector and each representation vector determined from the vector associated with an image containing a visual signature to be recognized, and
[0040] b3) return the identifier of the image containing a visual signature to be recognized for which the similarity value is the highest and exceeds a chosen threshold.
[0041] The invention also relates to a non-transient, computer-readable storage medium on which computer instructions are stored, in which the computer instructions are configured to cause a computer to execute a process as defined above.
[0042] The invention also relates to a computer program containing computer instructions which, when executed on a processor, cause the processor to execute the process as defined above.
[0043] Other features and advantages of the invention will become clearer from the following description, drawn from illustrative and non-limiting examples taken from the drawings in which: - Figure 1 is a schematic representation of an automatic recognition device according to the invention,
[0044] - Figure 2 shows an example of modules used in the computer program code of Figure 1, and
[0045] - Figure 3 is a flowchart of the process of recognizing a visual signature carried out by the device in Figure 1.
[0046] The drawings and description below contain, for the most part, elements of a definite nature. They can therefore not only serve to better explain the present invention, but also contribute to its definition, if necessary.
[0047] Figure 1 represents a general diagram of one embodiment of an automatic recognition device 100 according to the invention.
[0048] Device 100 is represented in Figure 1 as a functional diagram of a device (e.g., a network node, a connected device, etc.), according to one embodiment. As shown in Figure 1, this device may include: processing circuits (PC) 102, which may include one or more processors (P) or processing circuits 104; a network interface 106 comprising a transmitter (Tx) 108 and a receiver (Rx) 110 to enable the device to transmit data and receive data from other computing devices connected to a network 112 (e.g., an Internet Protocol (IP) network) to which the network interface 106 is connected; and data storage 114, which may include one or more non-volatile storage devices and / or one or more volatile storage devices.
[0049] In embodiments in which the PC 102 includes a programmable processor, a computer program product (CPP) 116 may be provided. The CPP 116 includes a computer-readable medium (CRM) 118 storing a computer program (CP) 120 comprising computer-readable instructions (CRI) 122. The CRM 118 may be a non-transient computer-readable medium, such as magnetic media (e.g., a hard disk drive), optical media, or memory devices (e.g., RAM, flash memory, etc.).The processor(s) or processing circuit(s) 104 include all known means for performing automated calculations, such as CPUs (Central Processing Unit), GPUs (Graphics Processing Unit), CPU and / or GPU grids, remote computing grids, specifically configured FPGAs (Field-Programmable Gate Array), specifically configured ASICs (Application-Specific Integrated Circuit), specialized chips such as SoCs (System on a Chip) or NOCs (Network on a Chip), AI-specialized chips, etc.
[0050] In the example described here, data storage 114 receives one or more input images and one or more visual signatures. One of the goals of device 100 is to determine whether the input image(s) contain one of the visual signatures, and where within the input image it may be located. As will be seen below, the input images can be received directly as inputs or obtained from input documents. The determination by device 100 involves detecting regions of interest, which will be described below and can also be stored temporarily or permanently in data storage 114. This determination also involves calculating associated vectors and representation vectors, which constitute a feature extraction of the input images and visual signatures. These elements can also be stored temporarily or permanently in data storage 114.
[0051] In the example described here, data storage 114 can be implemented in any suitable manner, i.e., by means of a hard drive, an SSD, flash memory, memory embedded in a processor, remote storage accessible in the cloud, etc. Data storage 114 can also store any transient data that may be generated during the execution of the invention, as well as data resulting from the operation of the device.
[0052] Figure 2 shows a schematic example of the computer program (CP) 120. The computer program (CP) 120 comprises two main modules which work together: a detector 212, an estimator 214. Optionally, the computer program 120 could also include a converter 216 upstream of the detector 212.
[0053] The 216 converter is conveniently designed to receive a document containing one or more pages and to return an image for each page of the document. The 216 converter is, for example, capable of receiving documents of any type and digital format.
[0054] The document can be any type of document converted into a digital image. The document can, for example, be taken from an online article, a scanned letter, a social media post, a patent document, an advertising poster, an invoice, a contract, a catalogue, a brochure, a photo, etc.
[0055] The image is a digital version of all or part of the document converted into a digital image, for example a capture of a page showing all or part of the document.
[0056] Detector 212 is configured to receive an image as input and to identify one or more regions of interest within it. In the examples described here, a region of interest is, for instance, a portion covering the image.
[0057] To perform this processing, detector 212 includes a convolutional neural network arranged to receive the image received by detector 212 as input, and to return as output one or more regions of interest if such a region is detected.
[0058] The convolutional neural network was trained to detect the potential presence of one or more visual signatures in images of the type shown. To do this, it returns regions of interest, each of which is a portion of the image.
[0059] The Applicant determined that, in order to improve the performance of estimator 214, it is preferable for the regions of interest to be bounding box images. For this reason, the Applicant chose to implement the convolutional neural network with a YOLO-type architecture or a faster R-CNN-type architecture.
[0060] In a classic way, a YOLO type neural network (for example Y0L0v8) includes a basic network (from the English backbone), an intermediate layer (from the English neck) and a final part (from the English head).
[0061] The base network is responsible for identifying image properties. In the example described here, the base network is of type cspdarknet53 (with 53 convolutional layers) which uses cross-stage partial connections. The Applicant found that using this type of base network improves information flow between the convolutional layers and increases accuracy. Alternatively, other base networks could be used, as is known for the various implementations of the YOLO architecture.
[0062] The intermediate layer is responsible for merging features extracted from the image at different stages by the core network. In the example described here, the intermediate layer is a C2f layer, enabling object detection within the image. Alternatively, other intermediate layers can be used, as is known for the various implementations of the YOLO architecture.
[0063] The final part is responsible for estimating regions of interest based on the work done by the core network and the intermediate layer. In the example described here, the image is fragmented into an array of N x N rectangles, where N is a natural number greater than or equal to two. One or more of these rectangles are grouped into a plurality of intermediate regions according to a probability that each rectangle belongs to a reference, such as an unwanted object reference or an object of interest reference. The membership probability is related to the accuracy of the YOLO-type network. Depending on the membership probability value, a class is assigned to each rectangle. A loss function is used during the training phase of the convolutional neural network to reduce potential discrepancies between the predicted and expected regions of interest during this training phase.Within the framework of the YOLO type architecture, the loss function is written according to the following equation:
[0064]
[0065] "
[0066] Or :
[0067] - L is the loss function for each rectangle,
[0068] ■ ^coord is a calculation difference between the predicted coordinate of the rectangle and the expected coordinate of the rectangle,
[0069] - L con f is a calculation difference between the predicted membership probability of the rectangle and the expected membership probability of the rectangle, and
[0070] - L cls is a difference in value between the predicted class and the expected class.
[0071] To improve training quality, the Applicant augmented the training data. To this end, one, some, or all of the images used for training were reproduced as modified images based on at least one modification parameter. This at least one image modification parameter was preferably chosen from the following group:
[0072] - a vertical flip of the image,
[0073] - a random translation of the image,
[0074] - a change in image depth,
[0075] - image scaling, and
[0076] - a combination of several input images.
[0077] Alternatively, the data increase might not be implemented.
[0078] As described above, the convolutional neural network can also be of the faster R-CNN type.
[0079] Typically, a faster R-CNN neural network comprises a base network and a classifier. In the example described here, the base network is responsible for identifying image properties. In this example, the base network is a ResNet-50 with a Feature Pyramid Network (FPN) that generates intermediate regions of a fixed size. Alternatively, a different pyramid base network, based for example on another ResNet or a DenseNet, could be used.
[0080] Downstream of the base network, a classifier categorizes an object detected in each of the intermediate regions into several references, such as the unwanted object reference and the object of interest reference. This time, each classified object has a probability of certainty (equivalent to the probability of each rectangle belonging to the YOLO network) linked to the accuracy of the faster R-CNN network.
[0081] For training purposes, the loss function used in the case of the R-CNN type architecture is written according to the following equation:
[0082]
[0083] OR :
[0084] - L is the loss function for each rectangle,
[0085] - L cls is a value representing the difference between the predicted class and the expected class,
[0086] - L reg is a calculation error in the box regression,
[0087] ■re g is the number of regressions performed,
[0088] - N cis is a number of classes of the classifier, and
[0089] - X is a weighting parameter that balances the contribution between the gap L reg and the value L cis .
[0090] To improve training quality, the Applicant augmented the training data. To this end, one, some, or all of the images used for training were reproduced as modified images based on at least one modification parameter. This at least one image modification parameter was preferably chosen from the following group:
[0091] - random cropping of annotated regions of interest in the training data, - insertion of a space around the training image or the annotated regions of interest in the training data, and
[0092] - a horizontal flip (left-right type for example) of the training images.
[0093] Alternatively, the data increase might not be implemented.
[0094] As we will see, the regions of interest returned by detector 212 are processed by estimator 214 to compare them to images containing visual signatures and determine if any of these regions of interest contain a visual signature. As described below, this comparison is performed by feature extraction, which relies on the implementation of a transformer. For this reason, both the regions of interest and the images containing visual signatures must have a chosen dimension, determined by the input characteristics of the transformer. Therefore, at the output of detector 212, or at the input of estimator 214, the regions of interest and the images containing visual signatures are processed to obtain suitable input dimensions for the transformer. This scaling is standard practice.
[0095] A visual signature is advantageously a logo or any type of image whose presence in the images is to be determined. For example, a visual signature could be a flourish, a handwritten signature, or a symbol. Device 100 can be configured to search not for a single visual signature, but for several. Advantageously, several visual signatures are then grouped into a visual signature group. Device 100 then processes the image(s) successively or in parallel with the visual signature group to indicate at output whether each image contains one or more of the visual signatures from the visual signature group, or none.
[0096] This determination is practically performed by estimator 214, as will now be described. In principle, estimator 214 operates by processing a region of interest each time to derive a representation vector, which can be viewed as feature extraction or dimensionality reduction. This same reduction is performed on images containing visual signatures.
[0097] Estimator 214 compares the representation vector of the region of interest with each of the representation vectors of the images containing the visual signatures and determines which image has the greatest similarity in its representation vector (optionally with a similarity floor) to define the visual signature considered to be within the region of interest. It can therefore be advantageous to pre-calculate the representation vectors of a set of desired visual signatures, since these are likely to be used regularly.
[0098] As input, estimator 214 is configured to receive a region of interest and one or more images containing a visual signature to be recognized. As output, estimator 214 is configured to return a value representing the presence or absence of a visual signature in the region of interest, as well as its identifier if applicable.
[0099] The Applicant has determined that the use of a triple network within estimator 214 is particularly effective in increasing the accuracy of visual signature detection.
[0100] The triple neural network comprises three subnetworks of the same architecture.
[0101] The three subnetworks have identical connection weights between their constituent layers, as well as an identical architecture. Each subnetwork includes, respectively, a transformer that produces an associated vector, and an optional multilayer perceptron network that produces a representation vector. When there is no multilayer perceptron network, the representation vector and the associated vector are identical.
[0102] The layers are normalization layers that act as a regularization mechanism for overfitting the three subnetworks. In the example described here, the transformer (respectively for each subnetwork) is an image representation transformer from a bidirectional encoder (Beit). The transformer is configured to receive the image corresponding to the region of interest detected in the image and to output the associated vector. The transformer produces the representation vector with a dimension chosen to be smaller than that of the associated vector. Alternatively, another transformer could be used instead of the BeiT, for example, an SWIN (see the article by Ze Liu et al. "SWIN Transformer: Hierarchical Vision Transformer using Shifted Window", https: / / arxiv.org / abs / 2103.14030).
[0103] Estimator 214 is designed to determine a representation vector (for each transformer) from a vector associated with the input region of interest. Estimator 214 is further designed to calculate the representative similarity value between this representation vector and each representation vector. Estimator 214 is also designed to return the identifier of the image containing a visual signature to be recognized for which the similarity value is highest and exceeds a chosen threshold.
[0104] The chosen threshold is, for example, a predefined value during a learning phase or before an inference phase of the triple neural network.
[0105] Optionally, estimator 214 includes a region multiplier configured to take the region of interest as input and return as output a plurality of cloned regions corresponding to at least two display parameters. Each display parameter is chosen from the group comprising:
[0106] - a horizontal flip of the region of interest,
[0107] - a random resized crop of the region of interest, and
[0108] - a random variation of the colors of the region of interest. The representation vector (respectively for each transformer) is a projection of the region of interest into a complex latent space. Each representation vector is determined from the vector associated with an image containing a visual signature to be recognized.
[0109] The complex latent space comprises a plurality of dimensions into which the region of interest can be projected. For example, the complex latent space has 256 dimensions. In this case, the representation vector has a dimension less than 256, preferably less than 128. In various versions, the dimensions may differ from the values indicated here.
[0110] During the training phase of the triple neural network, multiple refinement iterations are performed. For each of these iterations, the triple neural network selects a limited number of triplets of interest. The fact that the training phase is performed dynamically (online training) allows only the most relevant triplet(s) to be retained from among a plurality of reference image triplets during the training phase. Each reference image triplet includes one instance of a visual signature to be recognized, a positive image, and a negative image. For example, a set of reference images contains four reference image triplets. The triple neural network thus selects at least two of the most relevant triplets of interest from among the four reference image triplets at each refinement iteration.
[0111] An improvement function is used during the training phase of the triple neural network to reduce the discrepancies between the predicted associated vector and the expected associated vector. The improvement function is written according to the following equation:
[0112]
[0113] > where:
[0114] - O is the improvement function
[0115] - oc is a margin of error,
[0116] - A is the visual signature, - P is the positive image.
[0117] - N is the negative image, and
[0118] - f is a representation function.
[0119] The Applicant determined that it is particularly advantageous to use the triple neural network because it improves the detection quality of device 100 by promoting the generation of associated vectors that are closer when the images are closer, and further apart when the images are different.
[0120] Thus, it becomes possible to use the output vectors of the triple neural network during inference to compare them to associated vectors of visual signatures to determine if these are present in the regions of interest.
[0121] The positive image includes, for example, the visual signature to be recognized.
[0122] The negative image, for example, has a signature distinct from the visual signature. Advantageously, each negative image in the reference image triplet is selected according to the following optimization condition:
[0123] < <
[0124]
[0125] - oc is a margin of error,
[0126] - A is the visual signature,
[0127] - P is the positive image,
[0128] - N is the negative image, and
[0129] - f is a representation function.
[0130] This negative image selection technique is also called semi-hard negative, which selects a negative instance N that is closer to A than the positive instance P, but still incorrectly classified. This selection offers a good balance between discrimination difficulty and learning stability. It avoids the use of overly difficult examples, thus improving learning stability. Furthermore, it accelerates convergence by exploiting significant triplets and reduces the risk of overfitting. Alternatively, negatives could be chosen using hard negative.
[0131] The representation function is a function belonging to the triple neural network and explained below.
[0132] When the YOLO type network or the faster R-CNN type network did not return the region of interest, then it means that the image does not contain a visual signature.
[0133] Advantageously, the reference image set is also projected into the complex latent space during the training phase. This allows the reference images from among the set of reference images similar to the region of interest (positive images) to be projected onto dimensions close to that of the region of interest. Conversely, reference images from among the set of reference images very different from the region of interest will be projected onto dimensions far removed from that of the region of interest (negative images).
[0134] The multilayer perceptron network is arranged to receive the associated vector at the output of the transformer. The multilayer perceptron network comprises two fully connected layers (one for each multilayer perceptron network).
[0135] Each fully connected layer allows the extraction of a final representation vector (respectively for each fully connected layer from the representation vector received as input). More precisely, the representation function allows the associated vector to be input and the final representation vector to be obtained as output.
[0136] The final representation vector corresponds, for example, to a projection into a latent space simplified by the multilayer perceptron network. As mentioned above, in an embodiment where each network of the triple neural network comprises only a transformer, the output representation vector is identical to the associated vector produced by the transformer.
[0137] Device 100 is configured, using one or more input images, to:
[0138] - call detector 212 to extract, if necessary, at least one region of interest for each input image,
[0139] - call estimator 214 with each region of interest at the output of detector 212, and one or more images including a visual signature to be recognized.
[0140] The operation of device 100 according to the invention will now be described with reference to figure 3 representing an example of a flowchart of an automatic visual signature recognition process implemented by device 100.
[0141] Prior to the recognition of the visual signature, a conversion step 290 of each page of the document into the image is carried out by the converter 216. As mentioned previously, the converter 216 is optional, and the device 100 can directly receive one or more input images.
[0142] During step 300, device 100 detects, via detector 212, one or more regions of interest. If no region of interest is detected, device 100 moves on to the next image (taken from the document by converter 216 or received as input).
[0143] When a region of interest is detected in operation 300, device 100 determines in operation 310 the representative value of the presence of a visual signature to be recognized in the region of interest using estimator 214.
[0144] More precisely, the pre-trained triple neural network of estimator 214 receives as input an image corresponding to the region of interest determined in operation 300 and returns the associated vector as output. Estimator 214 then determines the representation vector corresponding to the region of interest as described above. In parallel, or subsequently, estimator 214 is used to determine the representation vector of one or more images containing a visual signature that is sought in the input images. Finally, estimator 214 calculates the distance between the representation vector of the region of interest and the representation vector determined for each visual signature. For example, the cosine distance or any other relevant distance can be used to assess the proximity of these representation vectors.Based on these distances, estimator 214 can determine if a visual signature has been detected in a region of interest, firstly by determining the visual signature associated with the distance indicating the greatest similarity of representation vectors, and optionally by comparing this distance to a threshold to validate that this similarity is sufficient in absolute terms.
[0145] When operation 300 returns multiple regions of interest, operation 310 includes calculating the distances to each region of interest. Therefore, it is advantageous to pre-calculate the representation vectors of the visual signatures we are trying to detect, as this saves time in the inference process.
[0146] The Applicant's tests revealed that Device 100 and its implementation method offer rapid analysis (less than one minute using standard computing hardware on a set of one hundred documents of varying types and page lengths), as well as very high accuracy (94.7% correct analyses on average compared to ground truth performed by a human considering all fields of analysis). Furthermore, Device 100 exhibits deterministic behavior (the same documents tested multiple times yield the same results) and is highly scalable.
Claims
Demands
1. Device (100) for the automatic recognition of at least one visual signature, the device (100) comprising: - a detector (212) arranged to receive an input image and to return, if necessary, at least one region of interest, the detector (212) comprising a convolutional neural network arranged to receive the input image as input and to return a portion of the input image as a region of interest, the convolutional neural network having been trained to detect the potential presence of a visual signature in the image it receives as input; - an estimator (214) arranged to receive an input region of interest and at least one image including a visual signature to be recognized, and to return a value representative of the presence of the visual signature to be recognized in the input region of interest, 1 ' estimator (214) comprising a triple neural network arranged to receive an image as input and to return an associated vector as output, the triple neural network comprising three subnetworks of the same architecture and having identical connection weights between layers within each respective subnetwork, which subnetworks comprise a transformer arranged to receive said image and to output the associated vector, the estimator (214) being arranged to determine a representation vector from the vector associated with the input region of interest and to calculate a representative value of the similarity between this representation vector and each representation vector determined from the vector associated with an image including a visual signature to be recognized, and to return the identifier of the image including a visual signature to be recognized for which the similarity value is the highest and exceeds a chosen threshold, the device (100) being arranged, based on one or more input images, to: - call the detector (212) to extract, if necessary, at least one region of interest for each input image, - call the estimator (214) with each region of interest at the output of the detector (212), and one or more images including a visual signature to be recognized.
2. Device according to claim 1, wherein the estimator (214) comprises one or more connected layers receiving the vector at the output of the transformer and producing a representation vector of a chosen dimension less than that of said vector.
3. Device according to claim 1 or 2, wherein the convolutional neural network is of the YOLO type or of the faster R-CNN type.
4. A device according to claim 3, wherein the convolutional neural network is of the YOLO type only, comprising a YOLO multiplier arranged to receive as input the input image and to return as output a plurality of cloned images corresponding to a combination of at least two image modification parameters, each image modification parameter being chosen from the group comprising: - a vertical flip of the input image, - a random translation of the input image, - a change in the depth of the input image, - scaling of the input image, and - a combination of several input images from among the plurality of input images.
5. A device according to claim 3, wherein the convolutional neural network is of the faster R-CNN type only, comprising an R-CNN multiplier arranged to receive as input the input image and to return as output a plurality of cloned images corresponding to a combination of at least two image transformation parameters, each modification parameter being chosen from the group comprising: - a random reframing of at least one region of interest, - the insertion of a space around the input image, - an insertion of a space around one among at least one region of interest.
6. A device according to any one of the preceding claims, wherein the transformer used in the subnetworks is of the BeiT or SWIN type.
7. A device according to any one of the preceding claims, wherein the estimator (214) comprises a region multiplier arranged to receive as input the region of interest and to return as output a plurality of cloned regions corresponding to at least two display parameters, each display parameter being selected from the group comprising: - a horizontal flip of the augmented input image, - a random resized crop of the enlarged input image, and - a random variation of the colors of the augmented input image.
8. A device according to any one of the preceding claims, wherein the learning phase is carried out on the internet, the learning phase comprising a plurality of refinement iterations of the triple neural network, the triple neural network selecting at each refinement iteration at least one triplet of interest that is the most relevant from among a plurality of reference image triplets, each reference image triplet comprising the visual signature, a positive image containing the visual signature to be recognized and a negative image containing a signature distinct from the chosen visual signature.
9. Device according to claim 8, wherein each negative image of the reference image triplet is selected according to the following optimization condition: < < where: - oc is a margin of error, - A is the visual signature to recognize. - P is the positive image, - N is the negative image, and - f is a representation function.
10. A method for automatically recognizing at least one visual signature comprising the following operations: a) receiving one or more input images, submitting at least one or more of the input images each as input to a detector (212) comprising a convolutional neural network arranged to return as a region of interest a portion of the input image, the convolutional neural network having been trained to detect the potential presence of a visual signature, and returning, where appropriate, at least one region of interest, b) for each region of interest of operation a) and for at least one or more images including a visual signature to be recognized, bl) determine an associated vector obtained by submitting it to a triple neural network arranged to receive an image as input and to return an associated vector as output, the triple neural network comprising three subnetworks of the same architecture and having identical connection weights between layers within each respective subnetwork, which subnetworks include a transformer arranged to receive said image and to output the associated vector, b2) determine a representation vector from the vector associated with the input region of interest and calculate a representative value of the similarity between this representation vector and each representation vector determined from the vector associated with an image containing a visual signature to be recognized, and b3) return the identifier of the image containing a visual signature to be recognized for which the similarity value is the highest and exceeds a chosen threshold.
11. A non-transient, computer-readable storage medium on which computer instructions are stored, wherein the computer instructions are configured to cause a computer to execute the process according to claim 10.
12. A computer program containing computer instructions which, when executed on a processor, cause the processor to perform the process according to claim 10.