Image recognition model training method and device, image recognition method and device

Abstract

The present disclosure provides an image recognition model training method and device, and an image recognition method and device. The image recognition model training device uses an image sample set to train a first deep learning model and a second deep learning model respectively; the output result of the first deep learning model is divided into a first non-noise label sample set and a first noise label sample set, and the output result of the second deep learning model is divided into a second non-noise label sample set and a second noise label sample set; the intersection of the first non-noise label sample set and the second non-noise label sample set is used as a clean label sample set; the first deep learning model is continued to be trained using the second non-noise label sample set, and the second deep learning model is continued to be trained using the first non-noise label sample set. The present disclosure can efficiently identify clean label samples and effectively avoid the occurrence of error accumulation.

Description

Technical Field

[0001] The present disclosure relates to the field of information processing, and in particular to an image recognition model training method and device, and an image recognition method and device. Background Art

[0002] Face recognition is one of the basic research problems in the field of pattern recognition and computer vision research. It has great practical application value and can be widely used in face access control, unmanned supermarkets, unmanned warehousing and logistics, intelligent robots, etc. In recent years, due to the development of big data and convolutional neural networks, face recognition technology has made significant progress.

[0003] At present, face recognition databases obtained from the Internet all have noisy labels. The existence of noisy labels will inevitably reduce the generalization ability of the model. In order to solve this problem, the existing technology predicts the noisy labels to select clean label samples. Summary of the invention

[0004] The inventors found through research that the existing noise label prediction method is inefficient and cannot achieve large-scale face recognition. In addition, there is error accumulation in the prediction process, which leads to model deviation.

[0005] To this end, the present disclosure provides a solution that can efficiently identify clean label samples while avoiding error accumulation.

[0006] According to a first aspect of an embodiment of the present disclosure, a method for training an image recognition model is provided, comprising: using an image sample set to respectively train a first deep learning model and a second deep learning model; dividing the output result of the first deep learning model into a first non-noise label sample set and a first noise label sample set, and dividing the output result of the second deep learning model into a second non-noise label sample set and a second noise label sample set according to a preset noise label ratio and a corresponding loss function value; taking the intersection of the first non-noise label sample set and the second non-noise label sample set as a clean label sample set; if the loss function value of the clean label sample does not meet the preset condition, continuing to train the first deep learning model using the second non-noise label sample set, and continuing to train the second deep learning model using the first non-noise label sample set; if the loss function value of the clean label sample meets the preset condition, ending the training.

[0007] In some embodiments, before continuing to train the first deep learning model using the second non-noise label sample set, the weights of the clean label samples are increased; before continuing to train the second deep learning model using the first non-noise label sample set, the weights of the clean label samples are increased.

[0008] In some embodiments, the number n of noise label samples is determined based on a preset noise label ratio and the number of training samples; the output results of the first deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function values ​​are used as the first noise label sample set; the output results of the second deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function value are used as the second noise label sample set.

[0009] In some embodiments, the preset noise label ratio increases as the number of iterations increases.

[0010] In some embodiments, the preset noise label ratio r t for Where t is the number of iterations, T k , r are preset parameters.

[0011] In some embodiments, the first deep learning model and the second deep learning model are convolutional neural network models.

[0012] In some embodiments, a cross entropy loss function is used to calculate the loss function value of the clean label sample, wherein in the cross entropy loss function, the weight of the non-clean label sample is reduced.

[0013] According to a second aspect of an embodiment of the present disclosure, an image recognition model training device is provided, comprising: a first training module, configured to use an image sample set to train a first deep learning model and a second deep learning model respectively; a set sorting module, configured to divide the output result of the first deep learning model into a first non-noise label sample set and a first noise label sample set, and divide the output result of the second deep learning model into a second non-noise label sample set and a second noise label sample set according to a preset noise label ratio and a corresponding loss function value; and use the intersection of the first non-noise label sample set and the second non-noise label sample set as a clean label sample set; an identification module, configured to detect whether the loss function value of the clean label sample meets a preset condition; if the loss function value of the clean label sample meets the preset condition, the training is terminated; and a second training module, configured to continue training the first deep learning model using the second non-noise label sample set, and continue training the second deep learning model using the first non-noise label sample set, if the loss function value of the clean label sample does not meet the preset condition.

[0014] According to a third aspect of an embodiment of the present disclosure, there is provided an image recognition model training device, comprising: a memory configured to store instructions; a processor coupled to the memory, the processor being configured to execute a method as involved in any of the above embodiments based on the instructions stored in the memory.

[0015] According to a fourth aspect of an embodiment of the present disclosure, an image recognition method is provided, comprising: extracting feature information of each image to be identified in a set of images to be identified; processing the feature information using a first deep learning model or a second deep learning model trained by the training method involved in any of the above embodiments to identify images with clean labels.

[0016] According to the fifth aspect of the embodiments of the present disclosure, an image recognition device is provided, comprising: a feature extraction module, configured to extract feature information of each image to be identified in a set of images to be identified; an image processing module, configured to process the feature information using a first deep learning model or a second deep learning model trained by the training method involved in any of the above embodiments, so as to identify an image with a clean label.

[0017] According to a sixth aspect of an embodiment of the present disclosure, there is provided an image recognition device, comprising: a memory configured to store instructions; a processor coupled to the memory, the processor being configured to execute a method as involved in any of the above embodiments based on the instructions stored in the memory.

[0018] According to a seventh aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and when the instructions are executed by a processor, the method involved in any of the above embodiments is implemented.

[0019] Other features and advantages of the present disclosure will become apparent from the following detailed description of exemplary embodiments of the present disclosure with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0021] Figure 1 A flowchart of an image recognition model training method according to an embodiment of the present disclosure;

[0022] Figure 2 A schematic diagram of an image recognition model training framework according to an embodiment of the present disclosure;

[0023] Figure 3 A schematic diagram of the structure of an image recognition model training device according to an embodiment of the present disclosure;

[0024] Figure 4 A schematic diagram of the structure of an image recognition model training device according to another embodiment of the present disclosure;

[0025] Figure 5 A flowchart of an image recognition method according to an embodiment of the present disclosure;

[0026] Figure 6 A schematic diagram of the structure of an image recognition device according to an embodiment of the present disclosure;

[0027] Figure 7 It is a structural schematic diagram of an image recognition device according to another embodiment of the present disclosure. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present disclosure and its application or use. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0029] Unless specifically stated otherwise, the relative arrangement of components and steps, the numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure.

[0030] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0031] Technologies, methods, and apparatus known to ordinary technicians in the relevant field may not be discussed in detail, but where appropriate, such technologies, methods, and apparatus should be considered part of the authorization specification.

[0032] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0033] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0034] Figure 1 The following is a flow chart of an image recognition model training method according to an embodiment of the present disclosure. In some embodiments, the following steps of the image recognition model training method are performed by an image recognition model training device.

[0035] In step 101, a first deep learning model and a second deep learning model are trained respectively using a set of image samples.

[0036] In some embodiments, the image sample set includes multiple face image samples.

[0037] In some embodiments, the first deep learning model and the second deep learning model are convolutional neural network models.

[0038] In step 102, according to the preset noise label ratio and the corresponding loss function value, the output results of the first deep learning model are divided into a first non-noise label sample set and a first noise label sample set, and the output results of the second deep learning model are divided into a second non-noise label sample set and a second noise label sample set.

[0039] It should be noted here that for the output results of the first deep learning model and the second deep learning model, the corresponding loss function value can be calculated according to the sample label. If the loss function value is large, it indicates that the sample is likely to be a sample with a noisy label. On the contrary, if the loss function value is small, it indicates that the sample is likely to be a sample with a clean label.

[0040] In some embodiments, the number of noise label samples n is determined according to a preset noise label ratio and the number of training samples. The output results of the first deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function value are used as the first noise label sample set. The output results of the second deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function value are used as the second noise label sample set.

[0041] For example, if the number of training samples is 100 and the preset noise label ratio is 2%, the number of noise label samples can be determined to be 2. Therefore, the output results of the first deep learning model are sorted according to the size of the corresponding loss function value, and the two output results with the largest loss function value are used as the first noise label sample set, and the remaining 98 output results are used as the first non-noise label sample set. Correspondingly, the output results of the second deep learning model are sorted according to the size of the corresponding loss function value, and the two output results with the largest loss function value are used as the second noise label sample set, and the remaining 98 output results are used as the second non-noise label sample set.

[0042] In some embodiments, the preset noise label ratio increases with the increase of the number of iterations. That is, more training samples are kept at the beginning of training. As the iterations proceed, the discard ratio of noise label samples is increased, thereby increasing the efficiency of training.

[0043] In some other embodiments, the noise label ratio r is preset t for Where t is the number of iterations, Tk , r is the preset parameter, and min is the minimum function.

[0044] For example, T k Take 10, r is 10%. In the first iteration, min{0.1r,r}=0.1r, that is, the preset noise label ratio is 1%. In the second iteration, min{0.2r,r}=0.2r, that is, the preset noise label ratio is 2%. And so on. In the 11th iteration, min{1.1r,r}=r, that is, the preset noise label ratio is 10%. That is, after a certain number of iterations, the preset noise label ratio is kept unchanged, thereby avoiding discarding too many training samples.

[0045] In step 103, the intersection of the first non-noise label sample set and the second non-noise label sample set is used as a clean label sample set.

[0046] It should be noted here that if the same sample is identified as a non-noise label sample by both the first deep learning model and the second deep learning model, the sample can be regarded as a clean label sample. Accordingly, the non-clean label samples in the first non-noise label sample set and the second non-noise label sample set are regarded as high-confidence clean label samples.

[0047] Therefore, the recognition results of the first deep learning model and the second deep learning model can be divided into three categories, namely, noise label samples, high-confidence trusted label samples and trusted label samples.

[0048] In step 104, if the loss function value of the clean label sample does not meet the preset conditions, the first deep learning model is continuously trained using the second non-noise label sample set, and the second deep learning model is continuously trained using the first non-noise label sample set.

[0049] It should be noted here that if each model is trained independently, errors will continue to accumulate as iterations proceed. In order to solve this problem, the present disclosure proposes to provide a first non-noise label sample set obtained by the first deep learning model to the second deep learning model, and to provide a second non-noise label sample set obtained by the second deep learning model to the first deep learning model. That is to say, during the training process, the high-confidence clean label samples determined by the first deep learning model and the high-confidence clean label samples determined by the second deep learning model are exchanged. In this way, each model can correct its own error accumulation with the help of other models.

[0050] In addition, during the iteration process, the determined noise label samples will be discarded, thereby effectively reducing the error accumulation caused by the noise label samples.

[0051] In some embodiments, before continuing to train the first deep learning model using the second set of non-noise label samples, the weight of the clean label samples is increased. Before continuing to train the second deep learning model using the first set of non-noise label samples, the weight of the clean label samples is increased. That is, during the iteration process, by increasing the weight of the clean label samples, the efficiency of model training can be effectively improved.

[0052] In step 105, if the loss function value of the clean label sample meets the preset condition, the training is terminated.

[0053] In some embodiments, a cross entropy loss function is used to calculate the loss function value of the clean label sample, wherein in the cross entropy loss function, the weight of the clean label sample is increased by reducing the weight of the non-clean label sample.

[0054] For example, the loss function value for a clean label sample is:

[0055]

[0056] Among them, s is the scale parameter, for example, s is 32. m is the interval value, and x is the sample feature. is the cosine value of the angle between the kth normalized classifier feature and feature x, and K is the total number of categories. g(μ)≥1 is a reweighted function, and the parameter μ is a non-negative value. Obviously, when μ=0, g(μ)=1. The specific expression is:

[0057]

[0058] In the image recognition model training method provided in the above embodiment of the present disclosure, a dual-model architecture is used for training, and during the iteration process, noisy label samples are discarded and high-confidence clean label samples are exchanged, thereby effectively reducing error accumulation and efficiently identifying clean label samples.

[0059] Figure 2 A schematic diagram of an image recognition model training framework according to an embodiment of the present disclosure.

[0060] like Figure 2As shown, the first deep learning model and the second deep learning model are trained respectively by using the image sample set. According to the preset noise label ratio and the corresponding loss function value, the output result of the first deep learning model is divided into a clean label sample set 21, a high confidence clean label sample set 22 and a noise sample label set 23. The output result of the second deep learning model is divided into a clean label sample set 21, a high confidence clean label sample set 24 and a noise sample label set 25. By discarding the noise sample label sets 23 and 25, the high confidence clean label sample sets 22 and 24 are exchanged, so that the first deep learning model is trained using the clean label sample set 21 and the high confidence clean label sample set 24, and the second deep learning model is trained using the clean label sample set 21 and the high confidence clean label sample set 22. The cycle is iterated until the loss function value of the clean label sample meets the predetermined condition.

[0061] Figure 3 FIG. 1 is a schematic diagram of the structure of an image recognition model training device according to an embodiment of the present disclosure. Figure 3 As shown, the training device includes a first training module 31 , a collection sorting module 32 , an identification module 33 and a second training module 34 .

[0062] The first training module 31 is configured to respectively train the first deep learning model and the second deep learning model using the image sample set.

[0063] In some embodiments, the image sample set includes multiple face image samples.

[0064] In some embodiments, the first deep learning model and the second deep learning model are convolutional neural network models.

[0065] The set sorting module 32 is configured to divide the output results of the first deep learning model into a first non-noise label sample set and a first noise label sample set according to a preset noise label ratio and a corresponding loss function value, and divide the output results of the second deep learning model into a second non-noise label sample set and a second noise label sample set; and take the intersection of the first non-noise label sample set and the second non-noise label sample set as a clean label sample set.

[0066] It should be noted here that for the output results of the first deep learning model and the second deep learning model, the corresponding loss function value can be calculated according to the sample label. If the loss function value is large, it indicates that the sample is likely to be a sample with a noisy label. On the contrary, if the loss function value is small, it indicates that the sample is likely to be a sample with a clean label.

[0067] In some embodiments, the number of noise label samples n is determined according to a preset noise label ratio and the number of training samples. The output results of the first deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function value are used as the first noise label sample set. The output results of the second deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function value are used as the second noise label sample set.

[0068] For example, if the number of training samples is 100 and the preset noise label ratio is 2%, the number of noise label samples can be determined to be 2. Therefore, the output results of the first deep learning model are sorted according to the size of the corresponding loss function value, and the two output results with the largest loss function value are used as the first noise label sample set, and the remaining 98 output results are used as the first non-noise label sample set. Correspondingly, the output results of the second deep learning model are sorted according to the size of the corresponding loss function value, and the two output results with the largest loss function value are used as the second noise label sample set, and the remaining 98 output results are used as the second non-noise label sample set.

[0069] In some embodiments, the preset noise label ratio increases with the increase of the number of iterations. That is, more training samples are kept at the beginning of training. As the iterations proceed, the discard ratio of noise label samples is increased, thereby increasing the efficiency of training.

[0070] In some other embodiments, the noise label ratio r is preset t for Where t is the number of iterations, T k , r is the preset parameter, and min is the minimum function.

[0071] For example, T k Take 10, r is 10%. In the first iteration, min{0.1r,r}=0.1r, that is, the preset noise label ratio is 1%. In the second iteration, min{0.2r,r}=0.2r, that is, the preset noise label ratio is 2%. And so on. In the 11th iteration, min{1.1r,r}=r, that is, the preset noise label ratio is 10%. That is, after a certain number of iterations, the preset noise label ratio is kept unchanged, thereby avoiding discarding too many training samples.

[0072] The identification module 33 is configured to detect whether the loss function value of the clean label sample meets the preset condition; if the loss function value of the clean label sample meets the preset condition, the training is terminated.

[0073] It should be noted here that if the same sample is identified as a non-noise label sample by both the first deep learning model and the second deep learning model, the sample can be regarded as a clean label sample. Accordingly, the non-clean label samples in the first non-noise label sample set and the second non-noise label sample set are regarded as high-confidence clean label samples.

[0074] Therefore, the recognition results of the first deep learning model and the second deep learning model can be divided into three categories, namely, noise label samples, high-confidence trusted label samples and trusted label samples.

[0075] The second training module 34 is configured to continue training the first deep learning model using the second non-noise label sample set if the loss function value of the clean label sample does not meet the preset conditions, and continue training the second deep learning model using the first non-noise label sample set.

[0076] It should be noted here that if each model is trained independently, errors will continue to accumulate as iterations proceed. In order to solve this problem, the present disclosure proposes to provide a first non-noise label sample set obtained by the first deep learning model to the second deep learning model, and to provide a second non-noise label sample set obtained by the second deep learning model to the first deep learning model. That is to say, during the training process, the high-confidence clean label samples determined by the first deep learning model and the high-confidence clean label samples determined by the second deep learning model are exchanged. In this way, each model can correct its own error accumulation with the help of other models.

[0077] In addition, during the iteration process, the determined noise label samples will be discarded, thereby effectively reducing the error accumulation caused by the noise label samples.

[0078] In some embodiments, before continuing to train the first deep learning model using the second set of non-noise label samples, the weight of the clean label samples is increased. Before continuing to train the second deep learning model using the first set of non-noise label samples, the weight of the clean label samples is increased. That is, during the iteration process, by increasing the weight of the clean label samples, the efficiency of model training can be effectively improved.

[0079] In some embodiments, a cross entropy loss function is used to calculate the loss function value of the clean label sample, wherein in the cross entropy loss function, the weight of the clean label sample is increased by reducing the weight of the non-clean label sample.

[0080] In the image recognition model training device provided in the above embodiment of the present disclosure, a dual-model architecture is used for training, and during the iteration process, noisy label samples are discarded and high-confidence clean label samples are exchanged, thereby effectively reducing error accumulation and efficiently identifying clean label samples.

[0081] Figure 4 FIG. 1 is a schematic diagram of the structure of an image recognition model training device according to another embodiment of the present disclosure. Figure 4 As shown, the training device includes a memory 41 and a processor 42 .

[0082] The memory 41 is used to store instructions. The processor 42 is coupled to the memory 41. The processor 42 is configured to execute the instructions stored in the memory to implement the following. Figure 1 The method of any one of the embodiments.

[0083] like Figure 4 As shown, the device also includes a communication interface 43 for information exchange with other devices. At the same time, the device also includes a bus 44, through which the processor 42, the communication interface 43, and the memory 41 communicate with each other.

[0084] The memory 41 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory. The memory 41 may also be a memory array. The memory 41 may also be divided into blocks, and the blocks may be combined into virtual volumes according to certain rules.

[0085] In addition, the processor 42 may be a central processing unit (CPU), or may be an application-specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present disclosure.

[0086] The present disclosure also relates to a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, which are executed by a processor to implement the following Figure 1 The method of any one of the embodiments.

[0087] Figure 5 The following is a flow chart of an image recognition method according to an embodiment of the present disclosure. In some embodiments, the following steps of the image recognition method are performed by an image recognition device.

[0088] In step 501, feature information of each image to be identified in a set of images to be identified is extracted.

[0089] In step 502, using Figure 1 The first deep learning model or the second deep learning model trained by the training method involved in any embodiment processes the feature information to identify images with clean labels.

[0090] By utilizing Figure 1In the training method involved in any embodiment, the first deep learning model or the second deep learning model can be used for image recognition. Therefore, the trained first deep learning model or the second deep learning model can be used to identify images with clean labels in the image set to be identified.

[0091] Figure 6 FIG. 1 is a schematic diagram of the structure of an image recognition device according to an embodiment of the present disclosure. Figure 6 As shown, the image recognition device includes a feature extraction module 61 and an image processing module 62 .

[0092] The feature extraction module 61 is configured to extract feature information of each image to be identified in the set of images to be identified.

[0093] The image processing module 62 is configured to utilize Figure 1 The first deep learning model or the second deep learning model trained by the training method involved in any embodiment processes the feature information to identify images with clean labels.

[0094] Figure 7 FIG. 1 is a schematic diagram of the structure of an image recognition device according to another embodiment of the present disclosure. Figure 7 As shown, the image recognition device includes a memory 71 , a processor 72 , a communication interface 73 and a bus 74 . Figure 7 and Figure 5 The difference is that in Figure 7 In the illustrated embodiment, the processor 72 is configured to execute instructions stored in the memory to implement the following Figure 5 The method of any one of the embodiments.

[0095] The present disclosure also relates to a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, which are executed by a processor to implement the following Figure 5 The method of any one of the embodiments.

[0096] In some embodiments, the functional unit module described above can be implemented as a general-purpose processor, a programmable logic controller (PLC), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components or any appropriate combination thereof for performing the functions described in the present disclosure.

[0097] By implementing the training method involved in the present disclosure, the first deep learning model or the second deep learning model can be used to perform corresponding face recognition detection.

[0098] Test 1:

[0099] First, we use the database CASIA-Clean-R with clean label samples as the training database, and artificially add noise labels to the training samples. Then we perform corresponding tests on databases such as LFW, CALFW, CPLFW, AgeDB, and CFP. The corresponding test results are shown in Table 1. In Table 1, "Detection" in the strategy column refers to the use of traditional methods to detect noise labels, "Exchange" refers to the scheme of exchanging high-confidence clean label samples between two models, and "Weighting" refers to the scheme of increasing the weight of clean labels during the exchange of high-confidence clean label samples.

[0100]

[0101]

[0102] Table 1

[0103] It can be seen from Table 1 that as the noise level increases, for example, from 0.1 to 0.3, the performance of the model tends to gradually decrease, but the effectiveness of the solution proposed in the present disclosure becomes more and more obvious.

[0104] Test 2:

[0105] We used two training sets, MS-Celeb-1M-R and VggFace2-R, from the real public face database with noisy labels, to train, tested on the corresponding test sets, and compared with the existing schemes. The corresponding results are shown in Tables 2 and 3 respectively.

[0106]

[0107]

[0108] Table 2

[0109] CALFW CPLFW AgeDB CFP Baseline 90.11 86.30 92.81 95.50 MentorNet 90.14 85.41 92.70 95.20 Decoupling 90.23 86.14 93.90 95.85 DualNet 90.26 85.01 93.33 95.05 Mutual Learning 90.08 86.00 93.35 95.51 Co-Teaching 89.90 85.05 92.05 95.05 Co-Teaching+ 89.43 85.23 92.50 95.41 This public scheme 91.06 87.31 94.05 95.87

[0110] Table 3

[0111] The results in Table 2 and Table 3 verify that the present invention can effectively reduce error accumulation and efficiently identify noisy label samples.

[0112] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0113] The description of the present disclosure is given for the purpose of illustration and description, and is not intended to be exhaustive or to limit the present disclosure to the disclosed form. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present disclosure, and to enable those of ordinary skill in the art to understand the present disclosure and thereby design various embodiments with various modifications suitable for specific uses.

Claims

1. A method for training an image recognition model, comprising: Using the image sample set to train the first deep learning model and the second deep learning model respectively; According to the preset noise label ratio and the corresponding loss function value, the output results of the first deep learning model are divided into a first non-noise label sample set and a first noise label sample set, and the output results of the second deep learning model are divided into a second non-noise label sample set and a second noise label sample set, wherein the number of noise label samples n is determined according to the preset noise label ratio and the number of training samples, the output results of the first deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function value are used as the first noise label sample set, the output results of the second deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function value are used as the second noise label sample set, and the preset noise label ratio increases with the increase of the number of iterations; The intersection of the first non-noise label sample set and the second non-noise label sample set is used as a clean label sample set; If the loss function value of the clean label sample does not meet the preset condition, the first deep learning model is further trained using the second non-noise label sample set, and the second deep learning model is further trained using the first non-noise label sample set; If the loss function value of the clean label sample meets the preset conditions, the training ends.

2. The method according to claim 1, wherein: Before continuing to train the first deep learning model using the second set of non-noise labeled samples, increasing the weight of the clean labeled samples; Before continuing to train the second deep learning model using the first set of non-noise labeled samples, the weights of the clean labeled samples are increased.

3. The method according to claim 1, wherein: The preset noise label ratio r t for ; Where t is the number of iterations, T k , r are preset parameters.

4. The method according to claim 1, wherein: The first deep learning model and the second deep learning model are convolutional neural network models.

5. The method according to any one of claims 1 to 4, wherein: The loss function value of the clean label sample is calculated using the cross entropy loss function, where the weight of the non-clean label sample is reduced in the cross entropy loss function.

6. An image recognition model training device, comprising: A first training module is configured to train the first deep learning model and the second deep learning model respectively using the image sample set; A set sorting module is configured to divide the output results of the first deep learning model into a first non-noise label sample set and a first noise label sample set, and divide the output results of the second deep learning model into a second non-noise label sample set and a second noise label sample set according to a preset noise label ratio and a corresponding loss function value; use the intersection of the first non-noise label sample set and the second non-noise label sample set as a clean label sample set, wherein the number of noise label samples n is determined according to the preset noise label ratio and the number of training samples, the output results of the first deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function value are used as the first noise label sample set, the output results of the second deep learning model are sorted according to the size of the corresponding loss function value, and the n output results with the largest loss function value are used as the second noise label sample set, and the preset noise label ratio increases with the increase of the number of iterations; The recognition module is configured to detect whether the loss function value of the clean label sample meets the preset condition; if the loss function value of the clean label sample meets the preset condition, the training is terminated; The second training module is configured to continue training the first deep learning model using the second non-noise label sample set if the loss function value of the clean label sample does not meet the preset conditions, and continue training the second deep learning model using the first non-noise label sample set.

7. An image recognition model training device, comprising: a memory configured to store instructions; A processor is coupled to the memory, and the processor is configured to execute the method according to any one of claims 1 to 5 based on instructions stored in the memory.

8. An image recognition method, comprising: Extracting feature information of each image to be identified in the set of images to be identified; The feature information is processed using a first deep learning model or a second deep learning model trained by the training method of any one of claims 1 to 5 to identify images with clean labels.

9. An image recognition device, comprising: A feature extraction module is configured to extract feature information of each image to be identified in the set of images to be identified; An image processing module is configured to process the feature information using a first deep learning model or a second deep learning model trained by the training method of any one of claims 1 to 5 to identify images with clean labels.

10. An image recognition device, comprising: a memory configured to store instructions; A processor is coupled to the memory, and the processor is configured to execute the method according to claim 8 based on instructions stored in the memory.

11. A computer-readable storage medium, wherein: The computer-readable storage medium stores computer instructions, and when the instructions are executed by a processor, the method according to any one of claims 1 to 5 and 8 is implemented.

Images