Model training method and device, article identification method and device, equipment and storage medium

Abstract

The embodiment of the invention discloses a model training method. The method comprises the steps of obtaining article information of a to-be-trained article; wherein the article information of the to-be-trained article represents a basic attribute of the to-be-trained article and a user experience condition of the to-be-trained article; determining a target article corresponding to the to-be-trained article based on the article information of the to-be-trained article; wherein the target article is used for identifying whether the to-be-trained article is forged or not; and performing model training on the initial counterfeit article identification model based on the article information of the to-be-trained article and the article information of the target article to obtain a target counterfeit article identification model. The embodiment of the invention further discloses an article identification method, a device corresponding to the model training method and the article identification method, equipment and a storage medium.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a model training, object identification method, device, equipment and storage medium. Background Art

[0002] Machine learning models have been widely used in various fields such as education, military, finance, and the Internet. Generally, an initial neural network model is used to learn and train the data of counterfeit items to obtain a target neural network model for identifying counterfeit items. However, during the implementation process, the inventors found that there are at least the following problems in the prior art: only focusing on the data features of the counterfeit items themselves, there is a problem of low model accuracy. Summary of the invention

[0003] In order to solve the above technical problems, the embodiments of the present application hope to provide a model training, object identification method, device, equipment and storage medium, which solves the problem that the prior art only focuses on the data characteristics of the counterfeit object itself and has a low model accuracy.

[0004] The technical solution of this application is implemented as follows:

[0005] A method for training a counterfeit object identification model, the method comprising:

[0006] Acquire item information of the item to be trained; wherein the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained;

[0007] Based on the object information of the object to be trained, determining a target object corresponding to the object to be trained; wherein the target object is used to identify whether the object to be trained is counterfeit;

[0008] The initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain a target counterfeit object identification model.

[0009] In the above solution, the step of determining the target object corresponding to the object to be trained based on the object information of the object to be trained includes:

[0010] By using the embedding layer in the initial counterfeit object identification model, the object information of the object to be trained is encoded using a low-dimensional vector to obtain the object features of the object to be trained;

[0011] The target item is determined from the item pool based on the item feature of the item to be trained and the item feature of each item in the item pool.

[0012] In the above solution, determining the target item from the item pool based on the item feature of the item to be trained and the item feature of each item in the item pool includes:

[0013] Determining, by a recall layer in the initial counterfeit object identification model, the similarity between the object to be trained and each object in the object pool based on the object feature of the object to be trained and the object feature of each object in the object pool;

[0014] The target item is determined from the item pool based on the similarity and the recall threshold.

[0015] In the above scheme, the initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain the target counterfeit object identification model, including:

[0016] The initial counterfeit object identification model is trained based on the object features of the object to be trained and the object features of the target object to obtain the target counterfeit object identification model.

[0017] In the above scheme, the initial counterfeit object identification model is trained based on the object features of the object to be trained and the object features of the target object to obtain the target counterfeit object identification model, including:

[0018] Performing feature crossover on the object features of the object to be trained and the object features of the target object through the feature crossover layer in the initial counterfeit object identification model to obtain crossover features;

[0019] The initial counterfeit object identification model is trained based on the cross features to obtain the target counterfeit object identification model.

[0020] In the above solution, the initial counterfeit object identification model is trained based on the cross-features to obtain the target counterfeit object identification model, including:

[0021] Performing feature optimization on the cross features to obtain key features;

[0022] The initial counterfeit object identification model is trained based on the key features to obtain the target counterfeit object identification model.

[0023] In the above solution, the cross-features are optimized to obtain key features, including:

[0024] Based on the forget gate, the input gate and the output gate, a feature optimization network in the initial counterfeit object identification model is constructed;

[0025] The feature optimization network is used to perform feature optimization on the cross features to obtain the key features.

[0026] In the above solution, the feature optimization network is used to optimize the cross-features to obtain the key features, including:

[0027] When the cross feature passes through the gate cell in the feature optimization network for feature optimization for the tth time, the implicit state of the t-1th gate cell and the tth cross feature in the cross feature are spliced ​​to obtain a spliced ​​feature;

[0028] Determine the to-be-recalled feature of the t-th gate cell through the forget gate in the feature optimization network based on the concatenated feature, the weight corresponding to the forget gate, and the constant corresponding to the forget gate;

[0029] Determine the cell state of the t-th gate cell through the input gate in the feature optimization network based on the feature to be recalled, the concatenated feature, the weight corresponding to the input gate, and the constant corresponding to the input gate;

[0030] Through the output gate in the feature optimization network, the hidden state of the t-th gate cell is determined based on the splicing feature, the weight corresponding to the output gate, the constant corresponding to the output gate and the cell state of the t-th gate cell, until the cross features have all passed through the gate cells in the feature optimization network, and the hidden state of the n-th gate cell is determined as the key feature.

[0031] A method for identifying an object, the method comprising:

[0032] Get item information of items to be processed;

[0033] Using a target counterfeit item identification model to identify the item information of the item to be processed, and determining whether the item to be processed is counterfeit;

[0034] The training process of the target counterfeit item identification model is as follows:

[0035] Acquire item information of the item to be trained; the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained;

[0036] Based on the object information of the object to be trained, determining a target object corresponding to the object to be trained; the target object is used to identify whether the object to be trained is counterfeit;

[0037] The initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain the target counterfeit object identification model.

[0038] A counterfeit identification model training device, the counterfeit identification model training device comprising:

[0039] A first acquisition unit is used to acquire item information of the item to be trained; wherein the item information of the item to be trained represents the basic attributes of the item to be trained and the user experience of the item to be trained;

[0040] A first processing unit, configured to determine a target object corresponding to the object to be trained based on the object information of the object to be trained; wherein the target object is used to identify whether the object to be trained is counterfeit;

[0041] The first processing unit is further used to perform model training on the initial counterfeit object identification model based on the object information of the object to be trained and the object information of the target object to obtain a target counterfeit object identification model.

[0042] An object identification device, the object identification device comprising:

[0043] A second acquiring unit, used to acquire item information of the item to be processed;

[0044] A second processing unit, configured to use the target counterfeit item identification model to identify the item information of the item to be processed, and determine whether the item to be processed is counterfeit;

[0045] The training process of the target counterfeit item identification model is as follows:

[0046] Acquire item information of the item to be trained; wherein the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained;

[0047] Based on the object information of the object to be trained, determining a target object corresponding to the object to be trained; wherein the target object is used to identify whether the object to be trained is counterfeit;

[0048] The initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain a target counterfeit object identification model.

[0049] A counterfeit identification model training device, the counterfeit identification model training device comprising: a first processor, a first memory and a first communication bus;

[0050] The first communication bus is used to realize the communication connection between the first processor and the first memory;

[0051] The first processor is used to execute the counterfeit object identification model training program in the first memory to implement the steps of the counterfeit object identification model training method mentioned above.

[0052] An article identification device, the article identification device comprising: a second processor, a second memory, and a second communication bus;

[0053] The second communication bus is used to realize the communication connection between the second processor and the second memory;

[0054] The second processor is used to execute the object identification program in the second memory to implement the steps of the above-mentioned object identification method.

[0055] A storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the forged object identification model training method or the object identification method.

[0056] The model training, object identification method, device, equipment and storage medium provided by the embodiments of the present application obtain the object information of the object to be trained, and determine the target object corresponding to the object to be trained based on the object information of the object to be trained, and the target object is used to identify whether the object to be trained is counterfeit, wherein the target object is used to identify whether the object to be trained is counterfeit, that is, the target object is the basis for identifying whether the object is counterfeit, and the object that can identify whether the object is counterfeit is a regular object, that is, the target object is the regular object corresponding to the object to be trained; based on the object information of the object to be trained and the object information of the target object, the initial counterfeit object identification model is trained to obtain the target counterfeit object identification model. Since the target counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object that can identify whether the object to be trained is counterfeit, during the training process, not only the object characteristics of the object to be trained itself can be learned, but also the object characteristics of the object to be trained can be learned. The implicit association features between the to-be-trained object and the target object, wherein the implicit association features between the to-be-trained object and the target object, that is, the implicit association features between the to-be-trained object and the regular object, are comprehensively considered during the training process, and the model accuracy of the target counterfeit object identification model can be improved; in this way, the target counterfeit object identification model is trained by the object information of the to-be-trained object and the object information of the target object that can identify whether the to-be-trained object is counterfeit, which can not only learn the object features of the to-be-trained object itself, but also learn the implicit association features between the to-be-trained object and the regular object, comprehensively consider the object features of the to-be-trained object itself and the implicit association features between the to-be-trained object and the regular object, improve the model accuracy of the target counterfeit object identification model, and overcome the problem that the prior art only focuses on the data features of the counterfeit object itself and has a low model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A flowchart of a method for training a counterfeit identification model provided in an embodiment of the present application;

[0058] Figure 2 A schematic diagram of a flow chart of an object identification method provided in an embodiment of the present application;

[0059] Figure 3 A flowchart of another method for training a counterfeit identification model provided in an embodiment of the present application;

[0060] Figure 4 A schematic diagram of a network architecture of an initial counterfeit item identification model in a counterfeit item identification model training method provided in an embodiment of the present application;

[0061] Figure 5A flowchart of another method for training a counterfeit identification model provided in an embodiment of the present application;

[0062] Figure 6 A schematic diagram of a process for determining genuine product data corresponding to training data in a counterfeit product identification model training method provided in an embodiment of the present application;

[0063] Figure 7 A schematic diagram of a structure for determining cross-features in a counterfeit identification model training method provided in an embodiment of the present application;

[0064] Figure 8 A schematic diagram of the structure of a feature optimization network in a counterfeit identification model training method provided in an embodiment of the present application;

[0065] Fig. 9 A schematic diagram of the structure of a confusion matrix used for model evaluation in a counterfeit identification model training method provided in an embodiment of the present application;

[0066] Fig.10 A schematic diagram of a model evaluation result in a counterfeit identification model training method provided in an embodiment of the present application;

[0067] Fig.11 A schematic diagram of a model evaluation result in another counterfeit identification model training method provided in an embodiment of the present application;

[0068] Fig.12 A schematic diagram of the structure of a counterfeit identification model training device provided in an embodiment of the present application;

[0069] Fig.13 A schematic diagram of the structure of an object identification device provided in an embodiment of the present application;

[0070] Fig.14 A schematic diagram of the structure of a counterfeit identification model training device provided in an embodiment of the present application;

[0071] Fig.15 A schematic diagram of the structure of an object identification device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0073] It should be understood that the "embodiments of the present application" or "the aforementioned embodiments" mentioned throughout the specification mean that specific features, structures or characteristics related to the embodiments are included in at least one embodiment of the present application. Therefore, "in the embodiments of the present application" or "in the aforementioned embodiments" appearing throughout the specification may not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the various embodiments of the present application, the size of the sequence numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application. The above-mentioned sequence numbers of the embodiments of the present application are for description only and do not represent the advantages and disadvantages of the embodiments.

[0074] Unless otherwise specified, the electronic device performs any step in the embodiments of the present application, and the processor of the electronic device may perform the step. It is also worth noting that the embodiments of the present application do not limit the order in which the electronic device performs the following steps. In addition, the methods used to process data in different embodiments may be the same method or different methods. It should also be noted that any step in the embodiments of the present application can be independently executed by the electronic device, that is, when the electronic device executes any step in the following embodiments, it may not rely on the execution of other steps.

[0075] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0076] The present application embodiment provides a method for training a counterfeit identification model. The method can be applied to a counterfeit identification model training device. Figure 1 As shown, the method comprises the following steps:

[0077] Step 101: Obtain item information of an item to be trained.

[0078] The item information of the item to be trained represents the basic attributes of the item to be trained and the user experience of the item to be trained.

[0079] In an embodiment of the present application, the items to be trained are items used for model training; in a feasible implementation, the items to be trained may include multiple counterfeit items and / or multiple regular items; counterfeit items are items counterfeited based on regular items, also known as fakes; regular items are also known as genuine items; or, the items to be trained may include only fakes. The item information of the items to be trained can characterize the basic attributes of the items to be trained, the user's evaluation information on the training items, the sales situation of the items to be trained, etc.; wherein, the user's evaluation of the training items can characterize the user experience of the items to be trained. In a feasible implementation, the item information of the items to be trained may include basic attributes such as detailed description, label price, preferential price, brand, and application scenario of the items to be trained, evaluation information such as user's rating and rating description of the training items, and sales situation such as purchase information and sales mode of the items to be trained.

[0080] In a feasible implementation method, item information of the items to be trained can be collected from the target platform to obtain the item information of the items to be trained; wherein the target platform can be a sharing platform, a shopping platform, an educational platform, etc., and further, the target platform can be in the form of an application (Appltcatton, App), a website, and a system, etc., which is not limited to the embodiments of the present application; for example, if the target platform is a shopping platform, then the items to be trained can be counterfeits on the shopping platform, and the item information of the items to be trained can include but is not limited to: the basic attributes of the counterfeit, the user's evaluation information on the counterfeit, and the purchase status of the counterfeit.

[0081] Step 102: Determine a target object corresponding to the object to be trained based on the object features of the object to be trained.

[0082] The target object is used to identify whether the object to be trained is counterfeit.

[0083] In the embodiment of the present application, the target object is used to identify whether the object to be trained is counterfeit, that is, the target object is the basis for identifying whether the object is counterfeit, and generally the object is identified through the regular object, so the target object is the regular object corresponding to the object to be trained; in this way, determining the target object corresponding to the object to be trained means determining the genuine object corresponding to the object to be trained. Since fakes are produced in imitation of genuine goods, each fake has one or more corresponding genuine goods, and there is a certain correlation between the fake and the corresponding genuine goods. The fake and the corresponding genuine goods will have some similar features, and there will also be some distinctive features, which are helpful to identify whether the object is counterfeit.

[0084] In a feasible implementation, in the process of training a target counterfeit object identification model using the object information of the object to be trained, the target object corresponding to the object to be trained can be recalled based on the object information of the object to be trained, so that subsequent model training can be performed based on the object information of the object to be trained and the object information of the target object, thereby improving the accuracy of the target counterfeit object identification model.

[0085] Step 103: Perform model training on the initial counterfeit object identification model based on the object information of the object to be trained and the object information of the target object to obtain a target counterfeit object identification model.

[0086] In the embodiment of the present application, the initial counterfeit identification model is a pre-set and untrained model; in a feasible implementation, the initial counterfeit identification model can be a Recall-Matched Neural Networks (RMNN) model. The target counterfeit identification model is a model obtained by training the initial counterfeit identification model based on the item information of the item to be trained and the item information of the target item, that is, the target counterfeit identification model is a trained model.

[0087] The initial counterfeit item identification model is trained based on the item information of the item to be trained and the item information of the target item used to identify whether the item to be trained is a counterfeit. Not only the item characteristics of the item to be trained itself are focused on, but also the implicit relationship between the item to be trained and the regular item is focused on. In this way, during the training process, not only the item characteristics of the item to be trained itself can be learned, but also the implicit association characteristics between the item to be trained and the regular item can be learned. The item characteristics of the item to be trained itself and the implicit association characteristics between the item to be trained and the regular item are comprehensively considered, thereby improving the model accuracy of the target counterfeit item identification model, thereby making the accuracy of counterfeit items identified based on the target counterfeit item identification model higher.

[0088] It should be noted that if the objects to be trained are only fakes, then the real objects corresponding to the fakes are determined based on the object information of the fakes, and the model training is performed based on the object information of the fakes and the object information of the corresponding real objects. In this case, the object characteristics of the fakes themselves can be learned, and the object characteristics of the real objects, as well as the implicit association characteristics between the fakes and the real objects, can improve the model accuracy of the target counterfeit object identification model; if the objects to be trained include fakes and real objects, then the real object corresponding to each object to be trained is determined based on the object information of each object to be trained, and the model training is performed based on the object information of each object to be trained. The model is trained based on the item information of the fake / genuine items and their corresponding authentic items. In this case, the item features of the fake / genuine items themselves can be learned, and the corresponding authentic items can also be used to better identify whether the items to be trained are fakes, thereby learning the implicit association features between fakes and authentic items. In summary, regardless of whether the items to be trained include only fakes or both fakes and authentic items, by determining their corresponding authentic items and training the model based on the item information of the items to be trained and their corresponding authentic items, the implicit association features between fakes and authentic items can be learned, and the model performance and model accuracy can be improved.

[0089] The counterfeit item identification model training method provided in the embodiment of the present application trains the target counterfeit item identification model through the item information of the to-be-trained item and the item information of the target item that can identify whether the to-be-trained item is counterfeit. It can not only learn the item characteristics of the to-be-trained item itself, but also learn the implicit association characteristics between the to-be-trained item and the regular item. It comprehensively considers the item characteristics of the to-be-trained item itself and the implicit association characteristics between the to-be-trained item and the regular item, improves the model accuracy of the target counterfeit item identification model, and overcomes the problem that the prior art only focuses on the data characteristics of the counterfeit item itself and has a low model accuracy.

[0090] Based on the above embodiments, the present application provides an object identification method, which is applied to an object identification device. Figure 2 As shown, the method comprises the following steps:

[0091] Step 201: Obtain item information of an item to be processed.

[0092] Among them, the items to be processed are the items that currently need to be identified.

[0093] In a feasible implementation, item information of the items to be processed may be collected from the target platform; the item information of the items to be processed may include but is not limited to: basic attributes of the items to be processed, user evaluation information of the items to be processed, and purchase status of the items to be processed.

[0094] Step 202: Use the target counterfeit item identification model to identify the item information of the item to be processed to determine whether the item to be processed is counterfeit.

[0095] Among them, the training process of the target counterfeit item identification model is as follows:

[0096] Acquire item information of the item to be trained; the item information of the item to be trained represents basic attributes of the item to be trained and user experience of the item to be trained; based on the item information of the item to be trained, determine the target item corresponding to the item to be trained; the target item is used to identify whether the item to be trained is counterfeit; based on the item information of the item to be trained and the item information of the target item, perform model training on an initial counterfeit item identification model to obtain a target counterfeit item identification model.

[0097] In the embodiment of the present application, the target counterfeit identification model can be a pre-trained model, so that the target counterfeit identification model can be directly used to identify the object. Specifically, the object information of the object to be processed can be directly input into the target counterfeit identification model for identification to determine whether the object to be processed is counterfeit (fake).

[0098] The object identification method provided in the embodiment of the present application has a target counterfeit object identification model that is trained by using the object information of the object to be trained and the object information of the target object that can identify whether the object to be trained is counterfeit. Therefore, the target counterfeit object identification model can not only learn the object characteristics of the object itself, but also learn the implicit association characteristics between the object and the regular object, and comprehensively considers the object characteristics of the object itself and the implicit association characteristics between its corresponding regular object. Therefore, the model accuracy of the target counterfeit object identification model is relatively high. Then, the target counterfeit object identification model is used to identify the object to be processed, and the identification result is more accurate, which overcomes the problem that the prior art only focuses on the data characteristics of the counterfeit object itself and has a low model accuracy.

[0099] Based on the above embodiments, the present invention provides a method for training a counterfeit identification model. Figure 3 As shown, the method comprises the following steps:

[0100] Step 301: The counterfeit object identification model training device obtains object information of an object to be trained.

[0101] The item information of the item to be trained represents the basic attributes of the item to be trained and the user experience of the item to be trained.

[0102] Step 302: The counterfeit object identification model training device encodes the object information of the object to be trained using a low-dimensional vector through the embedding layer in the initial counterfeit object identification model to obtain the object features of the object to be trained.

[0103] In the embodiment of the present application, the item feature of the to-be-trained item is a feature obtained by encoding the item information of the to-be-trained item using a low-dimensional vector. Since the item information of the to-be-trained item is encoded using a low-dimensional vector, the item feature of the to-be-trained item is also a feature vector of the to-be-trained item. Figure 4 As shown in the figure, the embedding layer in the initial counterfeit identification model can be used to encode the item information of the training item using a low-dimensional vector to obtain multiple item features such as item feature 1, item feature 2, and item feature 3. In the scenario of counterfeit identification, there are many item features corresponding to the item information, and each item feature will be recalled, so the amount of calculation is relatively large. In order to complete the recall efficiently, a low-dimensional vector is used here to encode the item information of the training item, retain the meaning of the item itself, make similar information closer in the same space, reduce the data dimension, and improve the subsequent processing rate.

[0104] In a feasible implementation, the item information of the item to be trained can be embedded from the high-dimensional space to the low-dimensional space to reduce the data dimension and improve the subsequent processing rate.

[0105] Step 303: The counterfeit object identification model training device determines a target object from the object pool based on the object feature of the object to be trained and the object feature of each object in the object pool.

[0106] The target object is used to identify whether the object to be trained is counterfeit.

[0107] In an embodiment of the present application, the item pool may store multiple genuine goods and the item features of each genuine good; in a feasible implementation method, multiple genuine goods of different categories may be collected in advance, and then the item information of these genuine goods may be encoded using low-dimensional vectors to obtain the item features of these genuine goods, and then these genuine goods and the item features of each genuine good may be stored in the item pool accordingly, so as to facilitate the recall of the item features of the genuine goods corresponding to the items to be trained during the training process.

[0108] In one achievable manner, Figure 4 As shown, the recall layer in the initial counterfeit item identification model can be used to determine the genuine item (i.e., the target item) corresponding to the item to be trained from the item pool based on the item features of the item to be trained and the item features of each item in the item pool, and then determine the item features of the target item.

[0109] Step 304: The counterfeit object identification model training device performs model training on the initial counterfeit object identification model based on the object features of the object to be trained and the object features of the target object to obtain a target counterfeit object identification model.

[0110] In the embodiment of the present application, the initial counterfeit object identification model is trained based on the object features of the object to be trained and the object features of the target object, that is, the model is trained based on the feature vector of the object to be trained and the feature vector of its corresponding genuine object to obtain the target counterfeit object identification model.

[0111] Wherein, step 304 can be implemented by the following steps:

[0112] Step 304a: The counterfeit object identification model training device performs feature crossover on the object features of the training object and the object features of the target object through the feature crossover layer in the initial counterfeit object identification model to obtain crossover features.

[0113] In the embodiment of the present application, the cross feature is a feature obtained by performing feature crossover on the item feature of the training item and the item feature of the target item, and the cross feature can be a feature vector. In a feasible implementation, a feature crossover strategy can be adopted to perform feature crossover on the item feature of the training item and the item feature of the target item to obtain the cross feature; the feature crossover strategy can be to perform L1 norm subtraction on the training data and its corresponding real product data, or perform inner product processing, or perform outer product processing to determine the difference between the training item and its corresponding real product; Figure 4 As shown, the feature cross layer in the initial counterfeit item identification model can be used to match the item features of the training item with the corresponding item features of the genuine item one by one, and the cross feature can be obtained through the feature cross strategy. This cross feature can be considered as the identification result of the training data based on the data, thereby enhancing the identification ability of the model.

[0114] During the training process, after passing the recall layer, each piece of data to be trained can get the corresponding real data (i.e., the basis data). The basis data is considered to be the most likely to identify fakes. After finding the basis data, a feature cross layer, also known as a matching layer, can be set to allow these basis data to "identify" fakes. In the matching layer, the training data (i.e., the item features of the item to be trained) and the basis data (i.e., the item features of the target item) are cross-featured to obtain cross-features, and the cross-features are used to represent the identification results of the training items by the basis data.

[0115] Step 304b: The counterfeit object identification model training device performs model training on the initial counterfeit object identification model based on the cross-features to obtain a target counterfeit object identification model.

[0116] In an embodiment of the present application, the item information of the item to be trained can be divided into a test set and a training set, and the training set is used for model training, and the test set is used for model testing. In the process of training, it is hoped to explore the hidden relationship between the data. Therefore, when dividing the training set and the test set, it is necessary to pay attention to the division of the region. K-fold cross validation can be used to divide the training set and the test set to reduce the impact of destroying the implicit relationship between the data due to the division.

[0117] There are fakes only when there are real ones. For a fake, the factors that determine whether it is a fake include: the item characteristics of the fake itself, and the difference characteristics between the fake and its corresponding real item. Obviously, there is a connection between the fake and the real. However, the traditional algorithm does not know what the connection is between the fake and the real, and how to connect; illustratively, the item information of a fake is input into the algorithm, and it is impossible to know which one is the corresponding real item, nor is it possible to know the identification criteria for actually identifying fakes. However, the embodiment of the present application designs a recall layer. During training, for each data to be trained (i.e., the item characteristics of the item to be trained), the real data (i.e., the item characteristics of the target item) that can best identify whether it is a fake is found, and it is used as the "basis data" for the fake, so as to learn the implicit relationship between fakes and real items in the process of training the model, thereby improving the accuracy of the model.

[0118] It should be noted that, for the description of the same steps and the same contents in this embodiment as those in other embodiments, reference can be made to the description in other embodiments and will not be repeated here.

[0119] The counterfeit item identification model training method provided in the embodiment of the present application trains the target counterfeit item identification model through the item information of the to-be-trained item and the item information of the target item that can identify whether the to-be-trained item is counterfeit. It can not only learn the item characteristics of the to-be-trained item itself, but also learn the implicit association characteristics between the to-be-trained item and the regular item. It comprehensively considers the item characteristics of the to-be-trained item itself and the implicit association characteristics between the to-be-trained item and the regular item, improves the model accuracy of the target counterfeit item identification model, and overcomes the problem that the prior art only focuses on the data characteristics of the counterfeit item itself and has a low model accuracy.

[0120] Based on the above embodiments, the present invention provides a method for training a counterfeit identification model. Figure 5 As shown, the method comprises the following steps:

[0121] Step 401: The counterfeit object identification model training device obtains object information of an object to be trained.

[0122] The item information of the item to be trained represents the basic attributes of the item to be trained and the user experience of the item to be trained.

[0123] Step 402: The counterfeit object identification model training device encodes the object information of the object to be trained using a low-dimensional vector through the embedding layer in the initial counterfeit object identification model to obtain the object features of the object to be trained.

[0124] Step 403: The counterfeit item identification model training device determines the similarity between the item to be trained and each item in the item pool through the recall layer in the initial counterfeit item identification model based on the item features of the item to be trained and the item features of each item in the item pool.

[0125] In the embodiment of the present application, similarity represents the degree of similarity between the object to be trained and each object in the object pool. In a feasible implementation, a distance calculation formula can be used to determine the similarity between the object to be trained and each object in the object pool based on the object features of the object to be trained and the object features of each object in the object pool; wherein the distance calculation formula can be the L2 norm, that is, the L2 norm is used to calculate the similarity between the object to be trained and each object in the object pool; the smaller the value calculated by the L2 norm, the stronger the similarity between the objects.

[0126] Step 404: The counterfeit item identification model training device determines the target item from the item pool based on the similarity and recall threshold.

[0127] The target object is used to identify whether the object to be trained is counterfeit.

[0128] In the embodiment of the present application, the recall threshold can be preset, and the recall threshold can be set according to actual business needs, which is not limited in the embodiment of the present application. Since the implicit connection between fakes and genuine goods is unknown, it is impossible to directly recall the data with the strongest implicit connection during the recall. Therefore, the recall threshold N can be appropriately increased to allow the data with strong implicit connection to be recalled as much as possible, that is, by increasing the recall threshold N, as much effective evidence data as possible can be found.

[0129] In a feasible implementation, the top N items in the item pool with the highest similarity ranking can be determined as target items (i.e., the real items corresponding to the training items), and then the item features of the target items (i.e., the real item data) can be determined. Figure 6 As shown in the figure, for each piece of data to be trained, the recall layer can find the N real items corresponding to the item to be trained from the M items in the item pool, and then determine the N real items.

[0130] In the scenario of counterfeit identification, since the identification of counterfeits depends on their corresponding genuine products, for each piece of data to be trained, the recall rules are involved to make the data with high implicit connections recalled as much as possible. For example, the l2 norm distance between each piece of data to be trained and all genuine data is calculated, and the genuine product that is "closer" to it is determined as the genuine object of this piece of data to be trained. In this way, by paying attention to the correlation between counterfeits and their corresponding genuine products, the characteristics of counterfeits can be more highlighted, thereby improving the identification accuracy of the model.

[0131] Step 405: The counterfeit object identification model training device performs feature crossover on the object features of the training object and the object features of the target object through the feature crossover layer in the initial counterfeit object identification model to obtain crossover features.

[0132] In the scenario of counterfeit identification, in step 404, for each piece of training data, its corresponding N pieces of real product data are found. By designing a feature crossover strategy, the item features of the training items and the corresponding item features of the real products are matched one by one to obtain crossover features, which can highlight the difference between the training data and the real product data (that is, the difference between real products and the difference between fake products and real products) through this design, thereby achieving the purpose of improving the algorithm performance.

[0133] Step 406: The counterfeit item identification model training device performs feature optimization on the cross features to obtain key features.

[0134] In the embodiment of the present application, the key feature is a feature obtained by optimizing the cross feature; the key feature may be an optimized feature vector. Figure 3 As shown in FIG. 1 , the cross features can be optimized through the feature optimization network in the initial counterfeit identification model to obtain the key features; specifically, Figure 7 As shown, the t-th feature in the training data can be cross-referenced with the t-th feature in the real data until all features are cross-referenced to obtain a cross-reference feature.

[0135] Wherein, step 406 can be implemented by the following steps:

[0136] Step 406a: The counterfeit object identification model training device constructs a feature optimization network in the initial counterfeit object identification model based on the forget gate, the input gate, and the output gate.

[0137] In the embodiment of the present application, the feature optimization network is used to optimize the cross features to remove unimportant features in the cross features, retain the truly meaningful features in the cross features, and improve the model performance. The forget gate is used to decide which features to discard and which features to retain; the input gate is used to update the cell state of the gate cell; and the output gate is used to determine the value of the next hidden state. Figure 8 As shown in the figure, for each data to be trained (1, 2, 3…, k), N cross features can be recalled, each cross feature corresponds to a gate cell, then the N cross features need to pass through N gate cells, and the gate cells include forget gate, input gate and output gate.

[0138] Step 406b: The counterfeit identification model training device uses a feature optimization network to optimize the cross features to obtain key features.

[0139] At the recall layer, in order to increase the model's ability to mine features between real and fake products, the number of "dependent data" to be recalled is increased to ensure that the real data corresponding to the items to be trained is recalled as much as possible. However, each time the recall threshold is increased, the number of cross-features will double, adding a large number of useless features, which has a certain negative impact on the performance of the algorithm. Therefore, in order to learn truly meaningful features that can characterize the correlation between real and fake products as much as possible, a feature optimization network is designed to optimize the cross-features to obtain key features, use truly meaningful features to identify fakes, and improve the performance of the model.

[0140] Wherein, step 406b can be implemented by the following steps:

[0141] Step A: When the cross feature passes through the gate cell in the feature optimization network for feature optimization for the tth time, the counterfeit identification model training device performs splicing processing on the implicit state of the t-1th gate cell and the tth cross feature in the cross feature to obtain a spliced ​​feature.

[0142] In the embodiment of the present application, the implicit state of the t-1th gate cell is the output feature of the t-1th gate cell; the spliced ​​feature is the feature obtained by splicing the implicit state of the t-1th gate cell and the tth cross feature in the cross feature. Figure 8 As shown in the figure, the implicit state of the first gate cell is the data to be trained. At this time, the data to be trained is concatenated with the first cross feature to obtain the concatenated feature. Each subsequent pass through the gate cell is concatenated with the implicit state of the previous gate cell (i.e., the output of the previous gate cell) and the tth cross feature to obtain the concatenated feature. During processing, each eigenvalue in the feature is processed, that is, each eigenvalue of the total implicit state of the t-1th gate cell and each value in the tth cross feature are concatenated respectively.

[0143] Step B: The counterfeit identification model training device determines the to-be-recalled features of the t-th gate cell through the forget gate in the feature optimization network based on the concatenated features, the weight corresponding to the forget gate, and the constant corresponding to the forget gate.

[0144] In the embodiment of the present application, the forget gate is used to discard and retain information. The feature to be recalled is the feature vector output by the forget gate. In a feasible implementation, the feature to be recalled can be determined based on the spliced ​​feature, the weight corresponding to the forget gate, and the constant corresponding to the forget gate through the forgetting function of the forget gate; specifically, the spliced ​​feature (i.e., the implicit state of the t-1th gate cell and the tth cross feature) is passed to the activation function (sigmoid) together, so that each feature value in the spliced ​​feature is between 0 and 1, and the closer to 0, the more useless the information is, and the closer to 1, the more useful information it is; the forgetting function of the forget gate can be: f t =σ(W f ·[h t-1 ,x t ]+b f ), σ() is the sigmoid function, W f is the weight of the forget gate, b f is the bias constant of the forget gate, [h t-1 ,x t ] is the hidden state h of the t-1th gate cell t-1 and the tth cross feature x t The stitching features obtained after the stitching process.

[0145] Step C: The counterfeit identification model training device determines the cell state of the t-th gate cell through the input gate in the feature optimization network based on the feature to be recalled, the concatenated feature, the weight corresponding to the input gate, and the constant corresponding to the input gate.

[0146] In the embodiment of the present application, the input gate is used to update the cell state of the gate cell in the feature optimization network, that is, to determine the cell state of the gate cell this time. In a feasible implementation, the update function of the input gate can be used to determine the feature to be updated based on the splicing feature, the update weight corresponding to the input gate, and the update constant corresponding to the forget gate; the adjustment function of the input gate can be used to determine the adjustment feature corresponding to the feature to be updated based on the splicing feature, the adjustment weight corresponding to the input gate, and the adjustment constant corresponding to the forget gate; the update strategy is used to determine the hidden state of the t-1th gate cell based on the feature to be recalled, the hidden state of the t-1th gate cell, the feature to be updated, and the adjustment feature.

[0147] Among them, the operation of determining the feature to be updated based on the concatenated feature, the update weight corresponding to the input gate and the update constant corresponding to the forget gate through the update function of the input gate can be: passing the concatenated feature (i.e., the implicit state of the t-1th gate cell and the tth cross feature) together to the sigmoid function, and determining the feature to be updated through 0-1. The update function of the input gate can be: t =σ(W i ·[h t-1 ,xt ]+b i ), W i is the weight of the input gate, b i is the update constant corresponding to the input gate.

[0148] The operation of determining the adjustment feature corresponding to the feature to be updated based on the concatenated feature, the adjustment weight corresponding to the input gate, and the adjustment constant corresponding to the forget gate through the adjustment function of the input gate can be: passing the concatenated feature (i.e., the implicit state of the t-1th gate cell and the tth cross feature) to the tanh function to create an adjustment feature vector. The adjustment function of the input gate can be: W C is the adjustment weight of the input gate, b i is the tuning constant corresponding to the input gate.

[0149] Adopting the update strategy, the operation of determining the hidden state of the t-th gate cell based on the feature to be recalled, the hidden state of the t-1th gate cell, the feature to be updated and the adjustment feature can be: performing operation processing on the feature to be recalled and the hidden state of the t-1th gate cell, performing operation processing on the feature to be updated and the adjustment feature, and adding the two values ​​after operation processing to obtain the hidden state of the t-th gate cell; wherein, the feature to be recalled and the hidden state of the t-1th gate cell can be processed by dot multiplication, and the feature to be updated and the adjustment feature can be processed by dot multiplication; the update strategy of the input gate can be expressed as: C t-1 is the hidden state of the t-1th gate cell.

[0150] Step D, the counterfeit identification model training device determines the hidden state of the t-th gate cell through the output gate in the feature optimization network based on the spliced ​​features, the weight corresponding to the output gate, the constant corresponding to the output gate and the cell state of the t-th gate cell, until the cross features have passed through the gate cells in the feature optimization network, and the hidden state of the n-th gate cell is determined as the key feature.

[0151] In an embodiment of the present application, the input gate is used to update the cell state of the gate cell in the feature optimization network, that is, to determine the cell state of the current gate cell. In a feasible implementation, the output function of the output gate can be used to determine the output feature based on the splicing feature, the weight corresponding to the output gate, and the constant corresponding to the output gate; the adjustment function of the output gate is used to determine the implicit state of the t-th gate cell based on the output feature and the hidden state of the t-th gate cell. Among them, the output function of the output gate is used to determine the output feature based on the splicing feature, the weight corresponding to the output gate, and the constant corresponding to the output gate. The operation can be: passing the splicing feature (that is, the implicit state of the t-1th gate cell and the tth cross feature) together to the sigmoid function, and determining the output feature through 0-1. The output function of the output gate can be: t =σ(W o ·[h t-1 ,x t ]+b o ), W o is the weight of the output gate, b o is the constant corresponding to the output gate. Using the adjustment function of the output gate, the operation of determining the implicit state of the t-th gate cell based on the output feature and the cell state of the t-th gate cell can be: passing the cell state of the t-th gate cell to the tanh function for processing, and performing calculations on the processed value and the output feature to obtain the implicit state of the t-th gate cell. The adjustment function of the output gate can be: t =o t ⊙tanh(C t ).

[0152] Step 407: The counterfeit item identification model training device performs model training on the initial counterfeit item identification model based on the key features to obtain a target counterfeit item identification model.

[0153] In the embodiments of the present application, Figure 8 As shown in the figure, the key features can be input into the fully connected hidden layer for identification, and the model can be trained by the back-propagation gradient descent method. Finally, the trained model can identify the implicit correlation features between fakes and real goods. The identification results are more accurate and can better determine whether the item is a fake.

[0154] In order to determine whether the target counterfeit identification model has better improvements compared to the network model in the prior art, and to deal with the problem of less counterfeit data, we do not pay attention to the identification of genuine goods, but pay more attention to whether the counterfeit can be accurately identified. Even in the case of 99% accuracy, it may be because the model is more biased towards the identification of non-counterfeit goods, and the identification of counterfeit goods is not enough - for example, 900 non-counterfeit goods may be correctly identified, while 100 counterfeit goods may fail to be identified, and the counterfeit identification rate is 0%, but the accuracy is 99%, which is not helpful for evaluating the target counterfeit identification model. Therefore, it can be used Fig. 9 The confusion matrix shown in the figure is used to judge the model performance of the target counterfeit identification model. After training with new data, more counterfeit data are accurately identified (predicted as positive), then the method is considered effective. Each column of the confusion matrix represents the predicted category, such as Fig. 9 The prediction categories in include positive and negative categories. The positive category can be represented by 0, indicating that the predicted category is real; the negative sample is represented by 1, indicating that the predicted category is fake; the value in each column indicates the number of real data predicted as this category, such as Fig. 9 The true (Reference) categories in include Positive and Negative; each row represents the true attribution category of the data.

[0155] like Fig.10 As shown in the figure, the model is trained using data with imbalanced data (i.e., less fake data). When the recall number is set to 0, that is, no recall, it is a traditional model. Fig.10 As can be seen from (a) in the figure, the traditional model has a high accuracy rate for identifying genuine goods, but the accuracy rate for identifying fake goods is only 0.44, and it has no good ability to identify goods for picking; Fig.10 (b) and Fig.10 In (c), the RMNN model with increased recall has significantly improved the accuracy of counterfeit identification by 6% and 9% respectively. That is, as the recall increases, the performance of the model has been significantly improved. Fig.11 In the model training, the data after improving the data imbalance phenomenon is used to train the model. Fig.11 From (a) in the figure, we can see that the accuracy of the traditional model in identifying fakes is 0.79. Fig.11 (b) The RMNN (when the recall threshold is set to 2) has an accuracy rate of 0.82 for counterfeit identification, which is an improvement of 3%. Therefore, when training the target counterfeit identification model, recalling the real data corresponding to the training items can improve the identification accuracy of the target counterfeit identification model.

[0156] The counterfeit identification model training method provided in the embodiment of the present application uses the embedding layer to process features to improve the processing speed of the model; adds a recall strategy, and determines the real data corresponding to the object to be trained by setting the recall condition and the recall threshold; in the feature cross-layer, the object to be trained and the corresponding real data are cross-featured to learn the item features of the object to be trained itself and the implicit association features between the object to be trained and the regular object, so that through the recall layer and the feature cross-layer, the model can not only focus on the features of the fake itself, but also focus on the features of the real corresponding to the fake and the implicit association features between the fake and the real, and add the cross-features to the training to improve the identification accuracy of the model. After obtaining the cross-features, it is necessary to design a suitable optimization strategy to optimize the mined features and the original features; specifically, by constructing a gate structure and designing an algorithm strategy for the real and fake features, the mined real and fake association features are allowed to enter the network learning together with the original features to improve the performance of the model. It should be noted that the counterfeit item identification model training method provided in the embodiment of the present application adopts a coarse recall method in the recall layer to recall multiple genuine product data corresponding to each data to be trained, and after cross-matching the data to be trained with the multiple genuine product data in the feature cross-layer, a fine screening method is used in the feature optimization layer to optimize the cross-features, so as to extract truly meaningful features and filter out useless features, thereby improving the processing rate and model performance.

[0157] The counterfeit item identification model training method provided in the embodiment of the present application can be applied not only to the identification of counterfeit items, but also to the identification of illegal stores and other application scenarios where there is a distinction between true and false. For example, when applied to illegal stores, the data to be trained are store features such as the trademark, product, and user evaluation information of each store. During the training process, the store features of the regular stores corresponding to each data to be trained are recalled, and the model is trained based on the data to be trained and the store features of the regular stores to obtain the illegal store identification model. It should be understood that the embodiment of the present application seeks to protect the recall of the item features of the corresponding regular items when identifying items that are distinguished between true and false, and the model can learn the distinguishing features between the true and false through the recalled item features of the regular items, thereby improving the model performance.

[0158] It should be noted that, for the description of the same steps and the same contents in this embodiment as those in other embodiments, reference can be made to the description in other embodiments and will not be repeated here.

[0159] The counterfeit item identification model training method provided in the embodiment of the present application trains the target counterfeit item identification model through the item information of the to-be-trained item and the item information of the target item that can identify whether the to-be-trained item is counterfeit. It can not only learn the item characteristics of the to-be-trained item itself, but also learn the implicit association characteristics between the to-be-trained item and the regular item. It comprehensively considers the item characteristics of the to-be-trained item itself and the implicit association characteristics between the to-be-trained item and the regular item, improves the model accuracy of the target counterfeit item identification model, and overcomes the problem that the prior art only focuses on the data characteristics of the counterfeit item itself and has a low model accuracy.

[0160] Based on the above embodiments, the present invention provides a counterfeit identification model training device, which can be applied to Figures 1 to 3 In the counterfeit identification model training method provided in the embodiment corresponding to 5, referring to Fig.12 As shown, the counterfeit identification model training device 5 may include:

[0161] A first acquisition unit 51 is used to acquire item information of the item to be trained; wherein the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained;

[0162] A first processing unit 52 is used to determine a target object corresponding to the object to be trained based on the object information of the object to be trained; wherein the target object is used to identify whether the object to be trained is counterfeit;

[0163] The first processing unit 52 is further used to perform model training on the initial counterfeit object identification model based on the object information of the object to be trained and the object information of the target object to obtain a target counterfeit object identification model.

[0164] In other embodiments of the present application, the first processing unit 52 is specifically configured to perform the following steps:

[0165] Through the embedding layer in the initial counterfeit object identification model, the item information of the to-be-trained item is encoded using a low-dimensional vector to obtain the item features of the to-be-trained item;

[0166] Based on the item feature of the to-be-trained item and the item feature of each item in the item pool, a target item is determined from the item pool.

[0167] In other embodiments of the present application, the first processing unit 52 is specifically configured to perform the following steps:

[0168] Determining the similarity between the object to be trained and each object in the object pool based on the object features of the object to be trained and the object features of each object in the object pool through the recall layer in the initial counterfeit object identification model;

[0169] Based on the similarity and recall threshold, the target item is determined from the item pool.

[0170] In other embodiments of the present application, the first processing unit 52 is specifically configured to perform the following steps:

[0171] The initial counterfeit object identification model is trained based on the object features of the object to be trained and the object features of the target object to obtain a target counterfeit object identification model.

[0172] In other embodiments of the present application, the first processing unit 52 is specifically configured to perform the following steps:

[0173] Through the feature cross layer in the initial counterfeit object identification model, the item features of the training item and the item features of the target item are cross-featured to obtain cross features;

[0174] The initial counterfeit object identification model is trained based on the cross features to obtain the target counterfeit object identification model.

[0175] In other embodiments of the present application, the first processing unit 52 is specifically configured to perform the following steps:

[0176] Perform feature optimization on cross features to obtain key features;

[0177] The initial counterfeit object identification model is trained based on the key features to obtain a target counterfeit object identification model.

[0178] In other embodiments of the present application, the first processing unit 52 is specifically configured to perform the following steps:

[0179] Based on the forget gate, input gate and output gate, a feature optimization network in the initial counterfeit identification model is constructed;

[0180] The feature optimization network is used to optimize the cross features and obtain the key features.

[0181] In other embodiments of the present application, the first processing unit 52 is specifically configured to perform the following steps:

[0182] When the cross feature passes through the gate cell in the feature optimization network for feature optimization for the tth time, the hidden state of the t-1th gate cell and the tth cross feature in the cross feature are spliced ​​to obtain the spliced ​​feature;

[0183] Through the forget gate in the feature optimization network, the features to be recalled of the t-th gate cell are determined based on the concatenated features, the weights corresponding to the forget gate, and the constants corresponding to the forget gate;

[0184] Through the input gate in the feature optimization network, the cell state of the t-th gate cell is determined based on the features to be recalled, the concatenated features, the weights corresponding to the input gate, and the constants corresponding to the input gate;

[0185] Through the output gate in the feature optimization network, the hidden state of the t-th gate cell is determined based on the splicing features, the weight corresponding to the output gate, the constant corresponding to the output gate and the cell state of the t-th gate cell, until the cross features have passed through the gate cells in the feature optimization network, and the hidden state of the n-th gate cell is determined as the key feature.

[0186] It should be noted that the specific description of the steps performed by each unit can be found in Figures 1 to 3 The counterfeit identification model training method provided in the embodiment corresponding to 5 is not repeated here.

[0187] The counterfeit item identification model training device provided in the embodiment of the present application trains a target counterfeit item identification model through the item information of the to-be-trained item and the item information of the target item that can identify whether the to-be-trained item is counterfeit. It can not only learn the item characteristics of the to-be-trained item itself, but also learn the implicit association characteristics between the to-be-trained item and the regular item. It comprehensively considers the item characteristics of the to-be-trained item itself and the implicit association characteristics between the to-be-trained item and the regular item, improves the model accuracy of the target counterfeit item identification model, and overcomes the problem that the prior art only focuses on the data characteristics of the counterfeit item itself and has a low model accuracy.

[0188] Based on the above embodiments, the present invention provides an object identification device, which can be applied to Figure 2 In the object identification method provided in the corresponding embodiment, refer to Fig.13 As shown, the object identification device 6 may include:

[0189] A second acquisition unit 61 is used to acquire item information of an item to be processed;

[0190] The second processing unit 62 is used to identify the item information of the item to be processed by using the target counterfeit item identification model to determine whether the item to be processed is counterfeit;

[0191] Among them, the training process of the target counterfeit item identification model is as follows:

[0192] Obtaining item information of the item to be trained; wherein the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained;

[0193] Based on the object information of the object to be trained, determining a target object corresponding to the object to be trained; wherein the target object is used to identify whether the object to be trained is counterfeit;

[0194] The initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain a target counterfeit object identification model.

[0195] It should be noted that the specific description of the steps performed by each unit can be found in Figure 2 The object identification method provided in the corresponding embodiment will not be described again here.

[0196] The object identification device provided in the embodiment of the present application has a target counterfeit object identification model that is trained by using the object information of the object to be trained and the object information of the target object that can identify whether the object to be trained is counterfeit. Therefore, the target counterfeit object identification model can not only learn the object characteristics of the object itself, but also learn the implicit association characteristics between the object and the regular object. It comprehensively considers the object characteristics of the object itself and the implicit association characteristics between its corresponding regular object. Therefore, the model accuracy of the target counterfeit object identification model is relatively high. Then, the target counterfeit object identification model is used to identify the object to be processed, and the identification result is more accurate, which overcomes the problem that the prior art only focuses on the data characteristics of the counterfeit object itself and has a low model accuracy.

[0197] Based on the above embodiments, the embodiments of the present application provide a counterfeit identification model training device, which can be applied to Figures 1 to 3 In the counterfeit identification model training method provided in the embodiment corresponding to 5, referring to Fig.14 As shown, the counterfeit identification model training device 7 may include: a first processor 71, a first memory 72 and a first communication bus 73, wherein:

[0198] The first communication bus 73 is used to realize the communication connection between the first processor 71 and the first memory 72;

[0199] The first processor 71 is used to execute the counterfeit identification model training program in the first memory 72 to implement the following steps:

[0200] Obtaining item information of the item to be trained; wherein the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained;

[0201] Based on the object information of the object to be trained, determining a target object corresponding to the object to be trained; wherein the target object is used to identify whether the object to be trained is counterfeit;

[0202] The initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain a target counterfeit object identification model.

[0203] In other embodiments of the present application, the first processor 71 is used to execute the counterfeit identification model training program in the first memory 72 based on the item information of the item to be trained, and determine the target item corresponding to the item to be trained, so as to implement the following steps:

[0204] Through the embedding layer in the initial counterfeit object identification model, the item information of the to-be-trained item is encoded using a low-dimensional vector to obtain the item features of the to-be-trained item;

[0205] Based on the item feature of the to-be-trained item and the item feature of each item in the item pool, a target item is determined from the item pool.

[0206] In other embodiments of the present application, the first processor 71 is used to execute the counterfeit identification model training program in the first memory 72, based on the item features of the item to be trained and the item features of each item in the item pool, to determine the target item from the item pool, so as to implement the following steps:

[0207] Determining the similarity between the object to be trained and each object in the object pool based on the object features of the object to be trained and the object features of each object in the object pool through the recall layer in the initial counterfeit object identification model;

[0208] Based on the similarity and recall threshold, the target item is determined from the item pool.

[0209] In other embodiments of the present application, the first processor 71 is used to execute the counterfeit identification model training program in the first memory 72 to perform model training on the initial counterfeit identification model based on the item information of the item to be trained and the item information of the target item to obtain the target counterfeit identification model, so as to implement the following steps:

[0210] The initial counterfeit object identification model is trained based on the object features of the object to be trained and the object features of the target object to obtain a target counterfeit object identification model.

[0211] In other embodiments of the present application, the first processor 71 is used to execute the counterfeit identification model training program in the first memory 72 to perform model training on the initial counterfeit identification model based on the object features of the object to be trained and the object features of the target object to obtain the target counterfeit identification model, so as to implement the following steps:

[0212] Through the feature cross layer in the initial counterfeit object identification model, the item features of the training item and the item features of the target item are cross-featured to obtain cross features;

[0213] The initial counterfeit object identification model is trained based on the cross features to obtain the target counterfeit object identification model.

[0214] In other embodiments of the present application, the first processor 71 is used to execute the counterfeit identification model training program in the first memory 72 to perform model training on the initial counterfeit identification model based on the cross-features to obtain the target counterfeit identification model, so as to implement the following steps:

[0215] Perform feature optimization on cross features to obtain key features;

[0216] The initial counterfeit object identification model is trained based on the key features to obtain a target counterfeit object identification model.

[0217] In other embodiments of the present application, the first processor 71 is used to execute the counterfeit identification model training program in the first memory 72 to optimize the cross features and obtain key features to implement the following steps:

[0218] Based on the forget gate, input gate and output gate, a feature optimization network in the initial counterfeit identification model is constructed;

[0219] The feature optimization network is used to optimize the cross features and obtain the key features.

[0220] In other embodiments of the present application, the first processor 71 is used to execute the counterfeit identification model training program in the first memory 72 to optimize the cross features using a feature optimization network to obtain key features, so as to implement the following steps:

[0221] When the cross feature passes through the gate cell in the feature optimization network for feature optimization for the tth time, the hidden state of the t-1th gate cell and the tth cross feature in the cross feature are spliced ​​to obtain the spliced ​​feature;

[0222] Through the forget gate in the feature optimization network, the features to be recalled of the t-th gate cell are determined based on the concatenated features, the weights corresponding to the forget gate, and the constants corresponding to the forget gate;

[0223] Through the input gate in the feature optimization network, the cell state of the t-th gate cell is determined based on the features to be recalled, the concatenated features, the weights corresponding to the input gate, and the constants corresponding to the input gate;

[0224] Through the output gate in the feature optimization network, the hidden state of the t-th gate cell is determined based on the splicing features, the weight corresponding to the output gate, the constant corresponding to the output gate and the cell state of the t-th gate cell, until the cross features have passed through the gate cells in the feature optimization network, and the hidden state of the n-th gate cell is determined as the key feature.

[0225] It should be noted that the specific description of the steps executed by the first processor 71 can be referred to Figures 1 to 3The counterfeit identification model training method provided in the embodiment corresponding to 5 is not repeated here.

[0226] The counterfeit item identification model training device provided in the embodiment of the present application trains a target counterfeit item identification model through the item information of the to-be-trained item and the item information of the target item that can identify whether the to-be-trained item is counterfeit. It can not only learn the item characteristics of the to-be-trained item itself, but also learn the implicit association characteristics between the to-be-trained item and the regular item. It comprehensively considers the item characteristics of the to-be-trained item itself and the implicit association characteristics between the to-be-trained item and the regular item, improves the model accuracy of the target counterfeit item identification model, and overcomes the problem that the prior art only focuses on the data characteristics of the counterfeit item itself and has a low model accuracy.

[0227] Based on the above embodiments, the embodiments of the present application provide an object identification device, which can be applied to Figure 2 In the object identification method provided in the corresponding embodiment, refer to Fig.15 As shown, the article identification device 8 may include: a second processor 81, a second memory 82, and a second communication bus 83, wherein:

[0228] The second communication bus 83 is used to realize the communication connection between the second processor 81 and the second memory 82;

[0229] The second processor 81 is used to execute the item identification program in the second memory 82 to implement the following steps:

[0230] Get item information of items to be processed;

[0231] Using a target counterfeit item identification model to identify the item information of the item to be processed, and determining whether the item to be processed is counterfeit;

[0232] The training process of the target counterfeit item identification model is as follows:

[0233] Acquire item information of the item to be trained; the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained;

[0234] Based on the object information of the object to be trained, determining a target object corresponding to the object to be trained; the target object is used to identify whether the object to be trained is counterfeit;

[0235] The initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain the target counterfeit object identification model.

[0236] It should be noted that the specific description of the steps executed by the second processor 81 can be referred to Figure 2 The object identification method provided in the corresponding embodiment will not be described again here.

[0237] In the object identification device provided in the embodiment of the present application, since the target counterfeit object identification model is trained by the object information of the object to be trained and the object information of the target object that can identify whether the object to be trained is counterfeit, the target counterfeit object identification model can not only learn the object characteristics of the object itself, but also learn the implicit association characteristics between the object and the regular object, and comprehensively considers the object characteristics of the object itself and the implicit association characteristics between its corresponding regular object. Therefore, the model accuracy of the target counterfeit object identification model is relatively high, and then the target counterfeit object identification model is used to identify the object to be processed, and the identification result is more accurate, which overcomes the problem that the prior art only focuses on the data characteristics of the counterfeit object itself and has a low model accuracy.

[0238] Based on the foregoing embodiments, the embodiments of the present application provide a storage medium, which stores one or more programs, and the one or more programs can be executed by one or more processors to implement Figures 1 to 3 and 5 corresponding to the steps of the counterfeit identification model training method provided in the embodiment, or Figure 2 The corresponding embodiments provide steps of the object identification method.

[0239] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of hardware embodiments, software embodiments, or embodiments in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.

[0240] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0241] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0242] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0243] The above description is only a preferred embodiment of the present application and is not intended to limit the protection scope of the present application.

Claims

1. A method for training a counterfeit identification model, characterized in that: The method comprises: Acquire item information of the item to be trained; wherein the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained; Based on the object information of the object to be trained, determining a target object corresponding to the object to be trained; wherein the target object is used to identify whether the object to be trained is counterfeit; The initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain a target counterfeit object identification model.

2. The method according to claim 1, characterized in that: The determining, based on the object information of the object to be trained, a target object corresponding to the object to be trained, includes: By using the embedding layer in the initial counterfeit object identification model, the object information of the object to be trained is encoded using a low-dimensional vector to obtain the object features of the object to be trained; The target item is determined from the item pool based on the item feature of the item to be trained and the item feature of each item in the item pool.

3. The method according to claim 2, characterized in that The determining the target item from the item pool based on the item feature of the item to be trained and the item feature of each item in the item pool includes: Determining, by a recall layer in the initial counterfeit object identification model, the similarity between the object to be trained and each object in the object pool based on the object feature of the object to be trained and the object feature of each object in the object pool; The target item is determined from the item pool based on the similarity and the recall threshold.

4. The method according to any one of claims 1 to 3, characterized in that The performing model training on the initial counterfeit object identification model based on the object information of the object to be trained and the object information of the target object to obtain the target counterfeit object identification model includes: The initial counterfeit object identification model is trained based on the object features of the object to be trained and the object features of the target object to obtain the target counterfeit object identification model.

5. The method according to claim 4, characterized in that The performing model training on the initial counterfeit object identification model based on the object features of the object to be trained and the object features of the target object to obtain the target counterfeit object identification model includes: Performing feature crossover on the object features of the object to be trained and the object features of the target object through the feature crossover layer in the initial counterfeit object identification model to obtain crossover features; The initial counterfeit object identification model is trained based on the cross features to obtain the target counterfeit object identification model.

6. The method according to claim 5, characterized in that The performing model training on the initial counterfeit object identification model based on the cross features to obtain the target counterfeit object identification model includes: Performing feature optimization on the cross features to obtain key features; The initial counterfeit object identification model is trained based on the key features to obtain the target counterfeit object identification model.

7. The method according to claim 6, characterized in that The step of optimizing the cross features to obtain key features includes: Based on the forget gate, the input gate and the output gate, a feature optimization network in the initial counterfeit object identification model is constructed; The feature optimization network is used to perform feature optimization on the cross features to obtain the key features.

8. The method according to claim 7, characterized in that The adopting the feature optimization network to perform feature optimization on the cross features to obtain the key features includes: When the cross feature passes through the gate cell in the feature optimization network for feature optimization for the tth time, the implicit state of the t-1th gate cell and the tth cross feature in the cross feature are spliced ​​to obtain a spliced ​​feature; Determine the to-be-recalled feature of the t-th gate cell through the forget gate in the feature optimization network based on the concatenated feature, the weight corresponding to the forget gate, and the constant corresponding to the forget gate; Determine the cell state of the t-th gate cell through the input gate in the feature optimization network based on the feature to be recalled, the concatenated feature, the weight corresponding to the input gate, and the constant corresponding to the input gate; Through the output gate in the feature optimization network, the hidden state of the t-th gate cell is determined based on the splicing feature, the weight corresponding to the output gate, the constant corresponding to the output gate and the cell state of the t-th gate cell, until the cross features have all passed through the gate cells in the feature optimization network, and the hidden state of the n-th gate cell is determined as the key feature.

9. A method for identifying an object, characterized in that: The method comprises: Get item information of items to be processed; Using a target counterfeit item identification model to identify the item information of the item to be processed, and determining whether the item to be processed is counterfeit; The training process of the target counterfeit item identification model is as follows: Acquire item information of the item to be trained; the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained; Based on the object information of the object to be trained, determining a target object corresponding to the object to be trained; the target object is used to identify whether the object to be trained is counterfeit; The initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain the target counterfeit object identification model.

10. A counterfeit identification model training device, characterized in that: The device comprises: A first acquisition unit is used to acquire item information of the item to be trained; wherein the item information of the item to be trained represents the basic attributes of the item to be trained and the user experience of the item to be trained; A first processing unit, configured to determine a target object corresponding to the object to be trained based on the object information of the object to be trained; wherein the target object is used to identify whether the object to be trained is counterfeit; The first processing unit is further used to perform model training on the initial counterfeit object identification model based on the object information of the object to be trained and the object information of the target object to obtain a target counterfeit object identification model.

11. An object identification device, characterized in that: The device comprises: A second acquiring unit, used to acquire item information of the item to be processed; A second processing unit, configured to use the target counterfeit item identification model to identify the item information of the item to be processed, and determine whether the item to be processed is counterfeit; The training process of the target counterfeit item identification model is as follows: Acquire item information of the item to be trained; wherein the item information of the item to be trained represents basic properties of the item to be trained and user experience of the item to be trained; Based on the object information of the object to be trained, determining a target object corresponding to the object to be trained; wherein the target object is used to identify whether the object to be trained is counterfeit; The initial counterfeit object identification model is trained based on the object information of the object to be trained and the object information of the target object to obtain the target counterfeit object identification model.

12. A counterfeit identification model training device, characterized in that: The counterfeit identification model training device comprises: a first processor, a first memory and a first communication bus; The first communication bus is used to realize the communication connection between the first processor and the first memory; The first processor is used to execute the counterfeit identification model training program in the first memory to implement the steps of the counterfeit identification model training method according to any one of claims 1 to 8.

13. An object identification device, characterized in that: The article identification device includes: a second processor, a second memory, and a second communication bus; The second communication bus is used to realize the communication connection between the second processor and the second memory; The second processor is used to execute the object identification program in the second memory to implement the steps of the object identification method as claimed in claim 9.

14. A storage medium, characterized in that: The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the counterfeit object identification model training method according to any one of claims 1 to 8, or the steps of the object identification method according to claim 9.

Images