METHOD AND DEVICE FOR PRODUCEING AN IMAGE CLASSIFICATION SYSTEM

DE602017086132T2Active Publication Date: 2024-11-13COMMISSARIAT A LENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
DE602017086132
Authority / Receiving Office
DE · DE
Patent Type
Patents
Current Assignee / Owner
Priority Date
2016-12-06
Filing Date
2017-12-01
Publication Date
2024-11-13
Estimated Expiration
2037-12-01

AI Technical Summary

Technical Problem

Existing image labeling systems face challenges in reusing specialized modules trained on specific labels for labeling images with different labels, leading to difficulties in achieving efficient labeling performance.

Method used

A method that combines specific and generic descriptive data from convolutional neural networks to create an efficient image labeling system, where the first module is trained on a first learning corpus and the second module is trained on a second learning corpus with generic labels, and both modules are used to provide descriptive data to a downstream module for improved labeling performance.

Benefits of technology

This approach enhances labeling performance by providing a more generic representation of images, allowing for efficient reuse of the learned features from the first learning corpus to label images with new labels, resulting in better labeling accuracy compared to using only specific or generic descriptive data alone.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader
Need to check novelty before this filing date? Find Prior Art

Description

[0001] The present invention relates to a method for obtaining an image labeling system, a corresponding computer program and device, and an image labeling system.

[0002] The invention applies more particularly to a method for obtaining a computer-implemented image labeling system, in which labels indicate the presence of objects in images, according to claim 1.

[0003] For example, the first article "From generic to specified deep representations for visual recognition" by H. Azizpour, A. Razavian, J. Sullivan, A. Maki and S. Carlsson, published in 2015 in Computer Vision and Pattern Recognition Workshops (CVPR-W), describes a transfer learning process (from the English "transfer-learning" ) .More specifically, this paper proposes to train the AlexNet convolutional neural network from a first training corpus to obtain a first image labeling module. This paper further proposes to use the output of the first fully connected neural layer, which is the sixth layer of the network, as descriptive data of an image.

[0004] Indeed, this particular layer represents, according to the article, a good compromise when the final task is not known. The descriptive data is then provided as input to a downstream module to carry out its automatic learning so that the system can label images based on labels that may be different from those of the first training corpus.

[0005] In a second article «In "Factors of transferability for a generic convnet representation" by H. Azizpour, A. Razavian, J. Sullivan, A. Maki, and S. Carlsson, published in 2015 in IEEE Transaction on Pattern Analysis and Machine Intelligence (PAMI), pages 1790-1802, the authors study the influence of the number of labels in the first training corpus on the performance of the image labeling system. To do this, they remove certain classes and the corresponding images (or random images). US 2016 / 117587 A1 (YAN ZHICHENG [US] ET AL), April 28, 2016, discloses image labeling by machine learning.

[0006] It may thus be desirable to provide a method for obtaining an image labeling program which makes it possible to improve the labeling performance of the image labeling system.

[0007] There is therefore provided a method of obtaining a computer-implemented image labeling system, in which labels indicate the presence of objects in images, according to claim 1.

[0008] By means of the invention, the first images are labeled in the second training corpus using second labels that are more generic than the first labels, since images that were previously associated with different first labels are, in the second training corpus, associated with the same second label. Thus, the first descriptive data provide a more generic representation of the image to be labeled than the second descriptive data. Surprisingly, it is by combining the first “specific” descriptive data and the second “generic” descriptive data as input to the downstream module that a high-performance image labeling system can be obtained.

[0009] Each of the first labeling module and the second labeling module comprises successive processing layers starting with a first processing layer, the first upstream module comprises one or more successive processing layers of the first module and the second upstream module comprises one or more successive processing layers of the second module.

[0010] Also optionally, the processing layer(s) of the first upstream module comprise the first processing layer of the first module and the processing layer(s) of the second upstream module comprise the first processing layer of the second module.

[0011] Each of the first labeling module and the second labeling module comprises a convolutional neural network comprising, as successive processing layers, convolutional layers and neural layers that follow the convolutional layers.

[0012] Also optionally, the first upstream module includes the convolutional layers and only part of the neural layers of the first module and the second upstream module includes the convolutional layers and only part of the neural layers of the second module.

[0013] Also optionally, obtaining the second learning corpus comprises, for each of at least part of the first labels, the determination, in a predefined tree of labels including in particular the first labels, of a common ancestor of this first label and of at least one other first label, the determined common ancestor being the replacement label of this first label.

[0014] The method comprises machine learning on a computer at least a portion of the downstream module from a third training corpus comprising third images associated with third labels, such that, when the first upstream module and the second upstream module receive as input one of the third images, the downstream module provides an output conforming to the third label associated with this third image in the third training corpus, the first upstream module and the second upstream module remaining unchanged during the training.

[0015] Optionally also, the downstream module comprises, on the one hand, a first block designed to receive as input the first descriptive data and the second descriptive data and to provide as output global descriptive data and, on the other hand, a second block designed to receive as input the global descriptive data and to provide as output a labeling, and the method further comprises: machine learning on a computer of the first block from a fourth training corpus comprising the first images associated with pairs of labels, the pair of labels associated with each first image comprising the first label associated with the first image in the first training corpus and the second label associated with the first image in the second training corpus, such that, when the first upstream module and the second upstream module receive as input one of the first images, the first block provides an output conforming to the pair of labels associated with this first image in the fourth training corpus, the first upstream module and the second upstream module remaining unchanged during the training, after the machine learning of the first block, machine learning on a computer of the second block from the third training corpus, such that,when the first upstream stage and the second upstream stage receive one of the third images as input, the downstream module provides an output conforming to the third label associated with this third image in the third training corpus, the first upstream module, the second upstream module and the first block remaining unchanged during training.

[0016] There is also provided a computer program downloadable from a communication network and / or recorded on a computer-readable medium and / or executable by a processor, characterized in that it comprises instructions for executing the steps of a method according to the invention, when said program is executed on a computer.

[0017] Also provided is a device for obtaining an image labeling system, designed to implement a method according to the invention.

[0018] Also provided is an image labeling system obtained by a computer-implemented method according to the invention.

[0019] The invention will be better understood with the aid of the following description, given solely by way of example and with reference to the appended drawings in which: there figure 1 illustrates the successive steps of a method for obtaining an image labeling system, according to a first embodiment of the invention, the figures 2 to 6 schematically represent the operations carried out during stages of the process of the figure 1 , there figure 7 illustrates the successive steps of a method for obtaining an image labeling system, according to a second embodiment of the invention, the figures 8 And 9 schematically represent the operations carried out during stages of the process of the figure 7 , there figure 10schematically represents a device for obtaining an image labeling system.

[0020] In the following description, labeling an image involves associating a score for that image with each of a plurality of predetermined labels.

[0021] Furthermore, in the following description, a module may be both a hardware module, i.e. a micro-programmed or micro-wired module in dedicated integrated circuits without the intervention of a computer program, and a software module intended to be executed by a processing unit of a computer. Alternatively, a module may comprise some hardware parts and other software parts.

[0022] In reference to the figures 1 to 7 , a first method 100 for designing an image labeling system S will now be described.

[0023] During a stage 102 (illustrated on the figure 2), a first image labeling module M 1 is obtained.

[0024] For this, the first module M 1 is trained by machine learning on a computer from a first training corpus CA 1 comprising first images I 1 associated with first labels L 1 .

[0025] During training, the first module M 1 is configured so that, when it receives one of the first images I 1 as input, it provides as output a labeling conforming to the first label L 1 associated with this first image I 1 in the first training corpus CA 1 .

[0026] In the example described, the first module M 1 comprises successive processing layers starting with a first processing layer. The first processing layer is intended to receive the image to be labeled. The output of each processing layer is provided as input to the next processing layer, except for the last processing layer which provides a labeling of the image to be labeled. Each processing layer comprises parameters which are adjusted during training.

[0027] The first module M 1 is a convolutional neural network. Thus, the first processing layers are convolutional CT 1 layers (five in the described example) and the following layers are fully connected CT 2 neural layers (three in the described example).

[0028] The first CA 1 training corpus is for example the ILSVRC (ImageNet Large Scale Visual Recognition Challenge) image database which is itself extracted from the ImageNet image database, labeled according to the Word Net hierarchy.

[0029] The labels used in the most common image databases, including for example ILSVRC, are generally very specific. As a result, the images are very visually consistent and probably facilitate machine learning. However, their use leads to obtaining a very specialized first module M 1, which makes it difficult to reuse it for labeling based on other labels.

[0030] This is why a second, more generic image labeling module M 2 is obtained as will be described later, when describing steps 104 and 106.

[0031] During a step 103, the first module M 1 is again trained, but partially this time, by automatic learning on a computer from a third training corpus CA 3 comprising third images I 3 associated with third labels L 3 . The automatic learning is then carried out according to the method of “ fine tuning " ( "fine-tuning" in English) by adjusting only part of the parameters of the first module M 1 . For example, only the parameters of one or more last neural layers CT 2 are adjusted. For example, only the parameters of the last neural layer are adjusted. The parameters of the other layers remain at their values ​​from the first automatic training of the first module M 1 from the first training corpus CA 1 .

[0032] During a step 104, a second training corpus CA 2 is obtained from the first training corpus CA 1 . The second training corpus CA 2 comprises the first images I 1 associated with second labels L 2 .

[0033] Obtaining the second training corpus CA 2 involves replacing, in the first training corpus CA 1 , each of at least some of the first labels L 1 with a replacement label, at least two first labels L 1 being replaced with the same replacement label. The second labels L 2 include the replacement labels and any first labels L 1 that have not been replaced. Thus, all the first images I 1 that were associated, in the first training corpus CA 1 , with the same first label L 1 that has been replaced are associated, in the second training corpus CA 2 , with the replacement label of this first label L 1 . Furthermore, since at least two first labels L 1 are replaced with the same replacement label, the number of second labels L 2 is less than the number of first labels L 1 . Thus, the replacement labels are labels with a more generic meaning than the first labels L 1 that they replace.For example, a replacement label for the first labels "Labrador" and "German Shepherd" could be the label "dog" or the label "animal". Thus the second L 2 labels are more generic than the first L 1 labels.

[0034] For example, obtaining the second learning corpus CA 2 comprises, for each of at least part of the first labels L 1 , the determination, in a predefined tree of labels including in particular the first labels L 1 , of a common ancestor to this first label L 1 and to at least one other first label L 1 , the determined common ancestor forming the replacement label of this first label L 1 .

[0035] For example, the second training corpus CA 2 is obtained using the following algorithm. + a first set K1 of labels is initialized to contain the first labels L 1 and a second set K2 is initialized to the empty set, + repeat : for each label ci of the first set K1: for each label cj of the first set K1 different from the label ci: • determine the lowest common ancestor in the tree between label ci and label cj, • suggest to a user to replace ci with the closest common ancestor. If the user accepts the replacement, add the closest common ancestor to the second set K2, associate it with the firsts images associated with the label ci and exit the loop on the cj, if the label ci has not been replaced, add the label ci to the second set K2 while keeping its associations to the first images, reduce the second set K2 to unique labels by grouping if necessary the first associated images I 1, if the second set K2 is identical to the first set K1, exit the "repeat" loop. Otherwise, initialize the first set K1 with the labels of the second set K2 and initialize the second set to the empty set.

[0036] For example, the second training corpus CA 2 is obtained automatically from the first training corpus CA 1 and the tree, without user intervention. For example, in the previous algorithm, the step of accepting or not accepting a replacement by the user is replaced by a step during which the closest common ancestor is automatically added to the second set K2 on the basis of one or more predefined criteria. For example, the predefined criterion or criteria include the criterion according to which the closest common ancestor belongs to a predetermined list of labels of the tree and the label ci does not belong thereto.

[0037] For example again, each of at least some of the first L 1 labels is replaced by the closest common ancestor of this first L 1 label and at least one other first L 1 label, this closest common ancestor being located above a predetermined level in the tree, the levels of the tree being increasing from the leaf labels to the root label. For example, to obtain this result, the predefined criterion(s) for the preceding algorithm include the criterion that the closest common ancestor is above a predefined level in the tree and the label ci is below this predefined level.

[0038] For example, for each first label L 1 , a user is offered all the ancestors of this first label L 1 in the tree to select the replacement label for this first label L 1 .

[0039] Thus, obtaining the second CA 2 training corpus requires very little manual work and can be automated in whole or in part. In addition, no obtaining of new labeled images is necessary. Thus, the tedious work of labeling new images is avoided.

[0040] During a stage 106 (illustrated on the figure 3 ), the second image labeling module M 2 is obtained.

[0041] For this, the second module M 2 is trained by machine learning on a computer from the second training corpus CA 2 .

[0042] During training, the second module M 2 is configured so that, when it receives one of the first images I 1 as input, it provides as output a labeling conforming to the second label L 2 associated with this first image I 1 in the second training corpus CA 2 .

[0043] The second module M 2 is also a convolutional neural network. Thus, the first processing layers CT 1 are convolutional layers (five in the described example) and the following layers are fully connected neural layers CT 2 (three in the described example).

[0044] More preferably, the second module M 2 is obtained independently of the learning of the first module M 1 carried out in step 102. In other words, the first module M 1 obtained following step 102 is not used to obtain the second module M2.

[0045] During a step 107, the second module M 2 is again trained, but partially this time, by machine learning on a computer from the third training corpus CA 3 . The machine learning is then carried out according to the method of “ fine tuning " ( "fine-tuning"in English) by adjusting only part of the parameters of the second module M 2 . For example, only the parameters of one or more last CT 2 neural layers are adjusted. For example, only the parameters of the last neural layer are adjusted. The parameters of the other layers remain at their values ​​from the first automatic training of the second module M 2 from the second training corpus CA 2 .

[0046] During a step 108, the system S is obtained, so that it comprises three modules MAm 1 , MAm 2 and MAv. These three modules comprise, on the one hand, a first upstream module MAm 1 and a second upstream module MAm 2 designed to each receive the same image to be labeled I and to respectively provide first descriptive data DD 1 and second descriptive data DD 2 of the image to be labeled I and, on the other hand, a downstream module MAv designed to receive the first descriptive data DD 1 and the second descriptive data DD 2 and to provide a labeling L of the image to be labeled I.

[0047] The system S is for example obtained in the form of a computer program comprising instructions implementing the functions of the modules described above, when said computer program is executed on a computer. This computer program could just as well be divided according to all possible combinations into one or more sub-programs. The functions performed could also be at least partly micro-programmed or micro-wired in dedicated integrated circuits. Thus, as a variant, the system S could be an electronic device composed solely of digital circuits (without a computer program) for performing the same functions.

[0048] Step 108 of obtaining the system S comprises, for example, the following steps 110 to 114.

[0049] During a stage 110 (illustrated on the figure 4 ), the first upstream module MAm 1 is obtained from at least part of the first module M 1 .

[0050] For example, the first upstream module MAm 1 comprises one or more successive processing layers CT 1 , CT 2 of the first module M 1 , preferably including the first processing layer.

[0051] In the example described, the first upstream module MAm 1 includes the convolutional layers CT 1 and only part of the neural layers CT 2 , for example all the neural layers CT 2 except the last one.

[0052] Furthermore, in the described example, the first upstream module MAm 1 has a normalization layer N 1 placed after the last imported layer of the first module M 1 . The normalization layer N 1 is designed to normalize the output of the last imported layer of the first module M 1 according to a predefined vector norm, for example according to the Euclidean norm. Thus, in the described example, the normalized output of the last imported layer of the first module M 1 forms the first descriptive data DD 1 .

[0053] During a stage 112 (illustrated on the Figure 5 ), the second upstream module MAm 2 is obtained from at least part of the second module M 2 .

[0054] For example, the second upstream module MAm 2 comprises one or more successive processing layers CT 1 , CT 2 of the second module M 2 , preferably including the first processing layer.

[0055] In the example described, the second upstream module MAm 2 includes the convolutional layers CT 1 and only part of the neural layers CT 2 , for example all the neural layers CT 2 except the last one.

[0056] Furthermore, in the described example, the second upstream module MAm 2 has a normalization layer N 2 placed after the last imported layer of the second module M 2 . The normalization layer N 2 is designed to normalize the output of the last imported layer of the second module M 2 according to a predefined vector norm, for example according to the Euclidean norm. Thus, in the described example, the normalized output of the last imported layer of the second module M 2 forms the second descriptive data DD 2 .

[0057] Thus, the first descriptive data DD 1 are supplemented by the second descriptive data DD 2 which allow the image to be described at a more generic level than the first descriptive data DD 1 alone. This allows, as will be described later, an efficient reuse of the machine learning carried out from the first training corpus CA 1 to label images according to new labels.

[0058] During a stage 114 (illustrated on the figure 6 ), the downstream module MAv is obtained.

[0059] The downstream module MAv is trained by machine learning on a computer from the third training corpus CA 3 .

[0060] During training, the downstream module MAv is configured so that, when the first upstream module MAm 1 and the second upstream module MAm 2 receive one of the third images I 3 as input, the downstream module MAv provides as output a labeling conforming to the third label L 3 associated with this third image I 3 in the third training corpus CA 3 . During training, the first upstream module MAm 1 and the second upstream module MAm 2 remain unchanged.

[0061] In the example described, the downstream module MAv comprises a neural network, for example comprising three layers of neurons.

[0062] The number of third images I 3 may be much smaller than the number of first images I 1 . For example, the number of third images I 3 may be equal to or smaller than 10% of the number of first images I 1 . Furthermore, the third images I 3 may represent completely different things from the first images I 1 , and the third labels L 3 may be different from the first labels L 1 and the second labels L 2 . However, thanks to the presence of the second descriptive data DD 2 , it was found that the image labeling system S thus obtained gave good results for labeling images according to the third labels L 3 , in any case results often better than using the first descriptive data DD 1 alone.

[0063] In reference to the figures 7 to 9 , a second method 700 for obtaining an image labeling program P will now be described.

[0064] Method 700 is identical to Method 100 except for the differences which will now be described.

[0065] In the example described, the downstream module MAv comprises, on the one hand, a first block B 1 designed to receive as input the first descriptive data DD 1 and the second descriptive data DD 2 and to provide as output global descriptive data DDG combining in the example described the first descriptive data DD 1 and the second descriptive data DD 2 and, on the other hand, a second block B 2 designed to receive as input the global descriptive data DDG and to provide as output a labeling from these global descriptive data DDG.

[0066] Furthermore, in the example described, step 114 of obtaining the downstream module MAv comprises the following steps 702 and 704.

[0067] During a stage 702 (illustrated on the figure 8), the first block B 1 is trained by machine learning on a computer from a fourth training corpus CA 4 comprising the first images I 1 associated with pairs of labels L 1 , L 2 . The pair of labels L 1 , L 2 of each first image I 1 comprises the first label L 1 associated with the first image I 1 in the first training corpus CA 1 and the second label L 2 associated with the first image I 1 in the second training corpus CA 2 .

[0068] During training, the first block B 1 is configured so that, when the first upstream module MAm 1 and the second upstream module MAm 2 receive as input one of the first images I 1 , the first block B 1 provides as output a double labeling (corresponding in the example described to the global descriptive data DDG) conforming to the pair of labels L 1 , L 2 associated with this first image I 1 in the fourth training corpus CA 4 . During training, the first upstream module MAm 1 and the second upstream module MAm 2 remain unchanged.

[0069] In the example described, the first block B 1 is a neural network, for example comprising three layers of neurons.

[0070] During a stage 704 (illustrated on the figure 9 ), the second block B 2 is trained by machine learning on a computer from the third training corpus CA 3 .

[0071] During training, the second block B 2 is configured so that, when the first upstream module MAm 1 and the second upstream module MAm 2 receive as input one of the third images I 3 , the second block B 2 provides as output a labeling conforming to the third label L 3 associated with this third image I 3 in the third training corpus CA 3 . During training, the first upstream module MAm 1 , the second upstream module MAm 2 and the first block B 1 remain unchanged.

[0072] In the example described, the second block B 2 is a neural network, for example comprising three layers of neurons.

[0073] In reference to the figure 10 , a device 1000 for obtaining an image labeling system S will now be described.

[0074] The device 1000 comprises, for example, a computer comprising a processing unit 1002 (comprising, for example, one or more processors) and a memory 1004 (comprising, for example, a RAM memory) for storing data files and computer programs. The memory 1004 comprises, in particular, a program 1006 comprising instructions for carrying out some or all of the steps of a method for obtaining an image labeling system S as described above, when said program 1006 is executed on the computer by the processing unit 1002.

[0075] The program 1006 could just as well be divided according to all possible combinations into one or more subroutines. The steps performed could also be at least partly microprogrammed or microwired in dedicated integrated circuits. Thus, as a variant, the computer implementing the processing unit 1002 could be replaced by an electronic device composed solely of digital circuits (without a computer program) for carrying out the same steps.

[0076] It is clear that the methods described above make it possible to obtain an image labeling system using at least part of a first image labeling module trained from “specific” labels, and at least part of a second image labeling module trained from “generic” labels, which makes it possible to obtain good labeling performance.

[0077] It should also be noted that the invention is not limited to the embodiments described above. It will indeed appear to those skilled in the art that various modifications can be made to the embodiments described above, in light of the teaching which has just been disclosed to them. In the following claims, the terms used should not be interpreted as limiting the claims to the embodiments set out in the present description, but should be interpreted to include all equivalents which the claims are intended to cover by virtue of their wording and the prediction of which is within the reach of those skilled in the art by applying their general knowledge to the implementation of the teaching which has just been disclosed to them.

Claims

1. A method (100; 700) for obtaining a system (S) for labelling images implemented by computer, wherein labels indicate the presence of objects in images, the method comprising: - obtaining a first module (M1) for labelling images that has been trained by machine learning on a computer on the basis of a first training corpus (CA1) comprising first images (I1) associated with first labels (L1), in such a way that, when the first labelling module (M1) receives, as an input, one of the first images (I1), the first labelling module (M1) provides an output consistent with the first label (L1) associated with this first image (I1) in the first training corpus (CA1), - obtaining (108) the system (S) for labelling images in such a way that it comprises: • a first upstream module (MAm1) designed to receive an image to be labelled (I) and to provide first descriptive data (DD,) of the image to be labelled (I), the first upstream module (MAm1) being obtained from at least a portion of the first labelling module (M1), • a downstream module (MAv) designed to provide a labelling (L) of the image to be labelled (I) on the basis of the first descriptive data (DD1), - obtaining (104) a second training corpus (CA2) comprising the first images (I1) associated with second labels (L2) by replacing, in the first training corpus (CA1), each of at least a portion of the first labels (L1) by a replacement label, at least two first labels (L1) being replaced by the same replacement label, the second labels (L2) comprising the replacement labels and the possible first labels (L1) that have not been replaced, - the machine learning, on a computer, of a second module (M2) for labelling images on the basis of the second training corpus (CA2), in such a way that, when the second labelling module (M2) receives, as an input, one of the first images (I1), the second labelling module (M2) provides an output consistent with the second label (L2) associated with this first image (I1) in the second training corpus (CA2), the system (S) for labelling images further comprising a second upstream module (MAm2) designed to receive the image to be labelled (I) and to provide second descriptive data (DD2) of the image to be labelled (I), the second upstream module (MAm2) being obtained on the basis of at least a portion of the second labelling module (M2), and the downstream module (MAv) being designed to provide a labelling (L) of the image to be labelled (I) on the basis of the first descriptive data (DD,) and the second descriptive data (DD2), each of the first labelling module (M1) and the second labelling module (M2) comprising successive processing layers (CT1, CT2) beginning with a first processing layer designed to receive the first images (I1), and each of the first labelling module (M1) and the second labelling module (M2) comprising a convolutional neural network including, as successive processing layers, convolutional layers (CT,) and neural layers (CT2) following the convolutional layers (CT1), the method being characterised in that it further comprises: - the machine learning, on a computer, of at least a portion of the downstream module (MAv) on the basis of a third training corpus (CA3) comprising third images (I3) associated with third labels (L3), in such a way that, when the first upstream module (MAm1) and the second upstream module (MAm2) receive, as an input, one of the third images (I3), the downstream module (MAv) provides an output consistent with the third label (L3) associated with this third image (I3) in the third training corpus (CA3), the first upstream module (MAm1) and the second upstream module (MAm2) remaining unchanged during the automatic learning of said at least part of the downstream module (MAv).

2. The method (100; 700) according to claim 1, wherein the first upstream module (MAm1) comprises one or more successive processing layers (CT1, CT2) of the first labelling module (M1) and wherein the second upstream module (MAm2) comprises one or more successive processing layers (CT1, CT2) of the second labelling module (M2).

3. The method (100; 700) according to claim 2, wherein the processing layer(s) of the first upstream module (MAm1) comprise the first processing layer of the first labelling module (M1) and wherein the processing layer(s) of the second upstream module (MAm2) comprise the first processing layer of the second labelling module (M2).

4. The method (100) according to anyone of claims 1 to 3, wherein the first upstream module (MAm1) comprises the convolutional layers (CT,) and only a portion of the neural layers (CT2) of the first labelling module (M1) and wherein the second upstream module (MAm2) comprises the convolutional layers (CT,) and only a portion of the neural layers (CT2) of the second labelling module (M2).

5. The method (100, 700) according to any one of claims 1 to 4, wherein obtaining (104) the second training corpus (CA2) comprises, for each of at least a portion of the first labels (L1), the determination, in a predefined tree of labels including in particular the first labels (L1), of an ancestor common to this first label (L1) and to at least another first label (L1), the common ancestor determined being the replacement label of this first label (L1).

6. The method (700) according to any one of claims 1 to 5, wherein the downstream module (MAv) comprises, on the one hand, a first block (B1) designed to receive, as an input, the first descriptive data (DD,) and the second descriptive data (DD2) and to provide, as an output, global descriptive data (DDG) and, on the other hand, a second block (B2) designed to receive, as an input, the global descriptive data (DDG) and to provide, as an output, a labelling (L) and further comprising: - the machine learning, on a computer, of the first block (B1) on the basis of a fourth training corpus (CA4) comprising the first images (I1) associated with pairs of labels (L1, L2), the pair of labels (L1, L2) associated with each first image (I1) comprising the first label (L1) associated with the first image (I1) in the first training corpus (CA1) and the second label (L2) associated with the first image (I1) in the second training corpus (CA2), in such a way that, when the first upstream module (MAm1) and the second upstream module (MAm2) receive, as an input, one of the first images (I1), the first block (B1) provides an output consistent with the pair of labels (L1, L2) associated with this first image (I1) in the fourth training corpus (CA4), the first upstream module (MAm1) and the second upstream module (MAm2) remaining unchanged during the learning, - after the machine learning of the first block (B1), the machine learning, on a computer, of the second block (B2) on the basis of the third training corpus (CA3), in such a way that, when the first upstream stage (MAm1) and the second upstream stage (MAm2) receive, as an input, one of the third images (I3), the downstream module (MAv) provides an output consistent with the third label (L3) associated with this third image (I3) in the third training corpus (CA3), the first upstream module (MAm1), the second upstream module (MAm2) and the first block (B1) remaining unchanged during the learning.

7. A computer program that can be downloaded from a communication network and / or is recorded on a medium readable by computer and / or can be executed by a processor, characterised in that it comprises instructions for the execution of the steps of a method according to any one of claims 1 to 6, when said program is executed on a computer.

8. A device (1000) for obtaining a system (S) for labelling images, designed to implement a method according to any one of claims 1 to 6.

9. A system (S) for labelling images obtained by a computer-implemented method according to any one of claims 1 to 6.