Attribute recognition pre-training model generation method, attribute recognition model generation method

Abstract

Embodiments of the present disclosure disclose an attribute recognition pre-training model generation method and an attribute recognition model generation method. A specific embodiment of the method comprises: obtaining first item information; inputting a first item image into an image feature extraction model to obtain a first image feature matrix, and inputting each pre-stored item attribute information into an attribute feature extraction model to obtain a plurality of first attribute feature vectors; inputting the plurality of first attribute feature vectors into a mask processing model to obtain a plurality of sets of masked attribute feature vectors; inputting the first image feature matrix and the plurality of sets of masked attribute feature vectors into a masked attribute information generation model to generate at least one masked attribute information; and training an initial attribute recognition model to obtain a trained attribute recognition model as an attribute recognition pre-training model. The embodiment is related to artificial intelligence and can obtain an attribute recognition pre-training model that generates more accurate attribute information.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, specifically to a method for generating a pre-trained model for attribute recognition and a method for generating an attribute recognition model. Background Technology

[0002] Currently, item attribute information serves as the basic item information for an item. Item attribute information plays a crucial role in various application scenarios. The typical method for identifying the attribute information of a target item is to input the image information of the target item into a pre-trained attribute recognition model to output the item attribute information. This attribute recognition model can be a model generated based on a pre-trained model related to the attribute.

[0003] However, the inventors discovered that when using the above method to identify item attribute information, the following technical problems often arise:

[0004] Attribute recognition models rely solely on image features to identify item attributes, resulting in inaccurate and poor recognition performance.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this disclosure propose a method for generating a pre-trained model for attribute recognition and a method for generating an attribute recognition model to solve the technical problems mentioned in the background section above.

[0008] In a first aspect, some embodiments of this disclosure provide a method for generating a pre-trained model for attribute recognition, comprising: acquiring first item information for a first target item, wherein the first item information includes: a first item image and multiple pre-stored item attribute information; inputting the first item image into an image feature extraction model included in an initial attribute recognition model to obtain a first image feature matrix, and inputting each of the multiple pre-stored item attribute information into an attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining multiple first attribute feature vectors; inputting the multiple first attribute feature vectors into a masking processing model included in the initial attribute recognition model to perform masking processing on at least one of the multiple first attribute feature vectors, thereby obtaining multiple masked attribute feature vectors; inputting the first image feature matrix and the multiple masked attribute feature vectors into a masked attribute information generation model included in the initial attribute recognition model to generate at least one masked attribute information; and training the initial attribute recognition model based on the at least one first attribute feature vector and the at least one masked attribute information to obtain a trained attribute recognition model, which serves as a pre-trained model for attribute recognition.

[0009] Optionally, the above-mentioned inputting the first image feature matrix and the plurality of masked attribute feature vectors into the mask attribute information generation model included in the initial attribute recognition model to generate at least one mask attribute information includes: performing a linear transformation on the first image feature matrix to generate a transformation vector set; performing vector fusion of the transformation vector set with the plurality of masked attribute feature vectors to obtain a first fusion vector; inputting the first fusion vector into the mask vector generation model included in the mask attribute information generation model to generate the at least one mask prediction vector; and inputting each of the at least one mask prediction vectors into the classification model included in the mask attribute information generation model to generate mask attribute information, thereby obtaining the at least one mask attribute information.

[0010] Optionally, training the initial attribute recognition model based on the at least one first attribute feature vector and the at least one mask attribute information to obtain a trained attribute recognition model as an attribute recognition pre-training model includes: obtaining at least one first attribute information corresponding to the at least one first attribute feature vector; generating a loss value for the first target item based on the at least one first attribute information and the at least one mask attribute information; and training the initial attribute recognition model based on the loss value to obtain a trained attribute recognition model as an attribute recognition pre-training model.

[0011] Secondly, some embodiments of this disclosure provide an attribute recognition pre-trained model generation apparatus, comprising: a first acquisition unit configured to acquire first item information for a first target item, wherein the first item information includes: a first item image and a plurality of pre-stored item attribute information; a first input unit configured to input the first item image into an image feature extraction model included in an initial attribute recognition model to obtain a first image feature matrix, and to input each of the plurality of pre-stored item attribute information into an attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining a plurality of first attribute feature vectors; and a second input unit configured to... The aforementioned plurality of first attribute feature vectors are input to the masking processing model included in the aforementioned initial attribute recognition model to perform masking processing on at least one of the plurality of first attribute feature vectors to obtain a plurality of masked attribute feature vectors; the third input unit is configured to input the aforementioned first image feature matrix and the aforementioned plurality of masked attribute feature vectors to the mask attribute information generation model included in the aforementioned initial attribute recognition model to generate at least one mask attribute information; the first training unit is configured to train the aforementioned initial attribute recognition model based on the aforementioned at least one first attribute feature vector and the aforementioned at least one mask attribute information to obtain a trained attribute recognition model, which serves as an attribute recognition pre-training model.

[0012] Optionally, the third input unit can be configured to: linearly transform the first image feature matrix to generate a transformation vector set; fuse the transformation vector set with the multiple masked attribute feature vectors to obtain a first fused vector; input the first fused vector into a mask vector generation model included in the mask attribute information generation model to generate at least one mask prediction vector; and input each of the at least one mask prediction vectors into a classification model included in the mask attribute information generation model to generate mask attribute information, thereby obtaining at least one mask attribute information.

[0013] Optionally, the first training unit can be configured to: acquire at least one first attribute information corresponding to the at least one first attribute feature vector; generate a loss value for the first target item based on the at least one first attribute information and the at least one mask attribute information; and train the initial attribute recognition model based on the loss value to obtain a trained attribute recognition model, which serves as a pre-trained attribute recognition model.

[0014] Thirdly, some embodiments of this disclosure provide a method for generating an attribute recognition model, including: acquiring an attribute recognition pre-trained model and item labeling information for a second target item; determining task information corresponding to a downstream task, wherein the downstream task is a task associated with item attributes; in response to determining that the task information is attribute recognition task information, performing a first model structure adjustment on the attribute recognition pre-trained model to generate an adjusted attribute recognition model; and training the adjusted attribute recognition model according to the item labeling information to obtain the attribute recognition model.

[0015] Optionally, the method further includes: in response to determining that the task information is not attribute recognition task information, adjusting the model structure of the attribute recognition pre-trained model according to the task information to generate an initial task model for the task information; and training the initial task model according to the item labeling information to obtain a trained task model.

[0016] Optionally, the above-mentioned adjustment of the first model structure of the attribute recognition pre-trained model to generate the adjusted attribute recognition model includes: removing the mask processing model included in the attribute recognition pre-trained model, and modifying the mask attribute information generation model included in the attribute recognition pre-trained model to obtain the adjusted attribute recognition model.

[0017] Optionally, the aforementioned item labeling information includes: a second item image of the second target item and multiple item labeling attribute information; and the above-mentioned training of the adjusted attribute recognition model based on the aforementioned item labeling information to obtain the attribute recognition model includes: inputting the aforementioned second item image into the image feature extraction model included in the initial adjusted attribute recognition model to obtain a second image feature matrix, and inputting each item labeling attribute information from the multiple item labeling attribute information into the attribute feature extraction model included in the adjusted attribute recognition model to generate a second attribute feature vector, thereby obtaining multiple second attribute feature vectors; and performing linear processing on the aforementioned second image feature matrix. The attributes undergo a change, resulting in a second change vector set. This second change vector set and the multiple second attribute feature vectors are then fused to generate a second fused vector. This second fused vector is input into the attribute association information generation model included in the initially adjusted attribute recognition model to generate attribute association feature vectors. These attribute association feature vectors are then input into the classification model included in the initially adjusted attribute recognition model to generate multiple predicted attribute information. Multiple true attribute information for the second target item is obtained. Based on the multiple predicted attribute information and the multiple true attribute information, the adjusted attribute recognition model is trained to obtain the attribute recognition model.

[0018] Fourthly, some embodiments of this disclosure provide an attribute recognition model generation apparatus, comprising: a second acquisition unit configured to acquire an attribute recognition pre-trained model and item labeling information for a second target item; a determination unit configured to determine task information corresponding to a downstream task, wherein the downstream task is a task associated with an item attribute; a model structure adjustment unit configured to, in response to determining that the task information is attribute recognition task information, perform a first model structure adjustment on the attribute recognition pre-trained model to generate an adjusted attribute recognition model; and a second training unit configured to train the adjusted attribute recognition model based on the item labeling information to obtain the attribute recognition model.

[0019] Optionally, the above apparatus further includes: in response to determining that the task information is not attribute recognition task information, adjusting the model structure of the attribute recognition pre-trained model according to the task information to generate an initial task model for the task information; and training the initial task model according to the item labeling information to obtain a trained task model.

[0020] Optionally, the model structure adjustment unit can be configured to: remove the masking processing model included in the above-mentioned attribute recognition pre-trained model, and modify the masking attribute information generation model included in the above-mentioned attribute recognition pre-trained model to obtain the adjusted attribute recognition model.

[0021] Optionally, the aforementioned item labeling information includes: a second item image of the second target item and multiple item labeling attribute information; and the second training unit can be configured to: input the second item image into an image feature extraction model included in the initially adjusted attribute recognition model to obtain a second image feature matrix; input each item labeling attribute information in the multiple item labeling attribute information into an attribute feature extraction model included in the adjusted attribute recognition model to generate a second attribute feature vector, thereby obtaining multiple second attribute feature vectors; perform a linear transformation on the second image feature matrix to obtain a second transformation vector set; perform vector fusion on the second transformation vector set and the multiple second attribute feature vectors to generate a second fusion vector; input the second fusion vector into an attribute association information generation model included in the initially adjusted attribute recognition model to generate an attribute association feature vector; input the attribute association feature vector into a classification model included in the initially adjusted attribute recognition model to generate predicted attribute information; obtain multiple true item attribute information for the second target item; and train the adjusted attribute recognition model based on the predicted attribute information and the multiple true item attribute information to obtain an attribute recognition model.

[0022] Fifthly, some embodiments of this disclosure provide a method for generating full item attribute information, including: acquiring third item information for a third target item; inputting the third item information into a pre-trained attribute recognition model to output full item attribute information for the third target item, wherein the attribute recognition model is generated by the attribute recognition pre-training model generation method and the attribute recognition model generation method of some embodiments of this disclosure.

[0023] Sixthly, some embodiments of this disclosure provide a full-item attribute information generation apparatus, including: a third acquisition unit configured to acquire third item information for a third target item; and a fourth input unit configured to input the third item information into a pre-trained attribute recognition model to output full item attribute information for the third target item, wherein the attribute recognition model is generated by the attribute recognition pre-training model generation method and the attribute recognition model generation method of some embodiments of this disclosure.

[0024] In a seventh aspect, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any of the implementations of the first, third, and fifth aspects.

[0025] Eighthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, it implements the method as described in any of the implementations of the first, third, and fifth aspects.

[0026] Ninthly, some embodiments of this disclosure provide a computer program product, including a computer program that, when executed by a processor, implements the methods described in any of the implementations of the first, third, and fifth aspects above.

[0027] The above embodiments of this disclosure have the following beneficial effects: the attribute recognition pre-training model generation method of some embodiments of this disclosure can generate an attribute recognition pre-training model with more accurate attribute information. Specifically, the reason why the attribute recognition pre-training model is not accurate enough in identifying attribute information is that the attribute recognition model only uses the image feature information of the item to identify the item attribute information, which has the problems of insufficient recognition accuracy and poor recognition effect. Based on this, the attribute recognition pre-training model generation method of some embodiments of this disclosure firstly obtains first item information for a first target item. The first item information includes: a first item image and multiple pre-stored item attribute information. Here, the first item information is obtained as the data basis for subsequent model training. In addition, for the attribute information annotation of the first target item, the problem of a large amount of data to be annotated can be greatly reduced by using multiple pre-stored item attribute information. Next, the first item image is input into the image feature extraction model included in the initial attribute recognition model to obtain a first image feature matrix, and each of the multiple pre-stored item attribute information is input into the attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining multiple first attribute feature vectors. Here, image feature information and attribute feature information are extracted for subsequent attribute recognition models to learn the correspondence between image features and attribute features, and the correspondence between attribute features themselves. Next, the aforementioned multiple first attribute feature vectors are input into the masking model included in the initial attribute recognition model to mask at least one of the multiple first attribute feature vectors, obtaining multiple masked attribute feature vectors. Masking at least one first attribute feature vector facilitates the subsequent determination of the correspondence between attribute features. Then, the aforementioned first image feature matrix and the aforementioned multiple masked attribute feature vectors are input into the masked attribute information generation model included in the initial attribute recognition model to generate at least one masked attribute information. Here, the masked attribute information generation model can effectively learn not only the association between the first image feature matrix and each masked attribute feature vector, but also the attribute association between the various masked attribute feature vectors. Finally, based on the aforementioned at least one first attribute feature vector and the aforementioned at least one masked attribute information, the initial attribute recognition model is trained to obtain a trained attribute recognition model, which serves as a pre-trained model for attribute recognition. The resulting attribute recognition pre-trained model can effectively and accurately identify item attribute information by considering the correspondence between image features and attribute features, as well as the correspondence between attribute features themselves. In summary, the attribute recognition model, during training, not only considers the correspondence between image features and attribute features, but also fully considers the correspondence between attribute features themselves.Therefore, a more accurate attribute recognition pre-trained model can be obtained by generating attribute information. Attached Figure Description

[0028] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0029] Figure 1 This is a schematic diagram of an application scenario of the attribute recognition pre-trained model generation method according to some embodiments of the present disclosure;

[0030] Figure 2 This is a flowchart of some embodiments of the attribute recognition pre-trained model generation method according to this disclosure;

[0031] Figure 3 These are flowcharts of other embodiments of the attribute recognition pre-trained model generation method according to this disclosure;

[0032] Figure 4 This is a flowchart of some embodiments of the attribute recognition model generation method according to this disclosure;

[0033] Figure 5 This is a flowchart of some embodiments of the method for generating full item attribute information according to this disclosure;

[0034] Figure 6 These are schematic diagrams illustrating the structure of some embodiments of the attribute recognition pre-trained model generation apparatus according to this disclosure;

[0035] Figure 7 These are schematic diagrams illustrating the structure of some embodiments of the attribute recognition model generation apparatus according to this disclosure;

[0036] Figure 8 This is a schematic diagram of the structure of some embodiments of the full-item attribute information generation device according to the present disclosure;

[0037] Figure 9 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0038] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0039] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0040] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0041] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0042] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0043] Before performing any of the operations related to the collection, storage, and use of item information (such as item images and item attribute information) involved in this disclosure, the relevant organizations or individuals shall fulfill their obligations, including conducting an item information security impact assessment, informing the item information subject, and obtaining the item information subject's prior authorization and consent.

[0044] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0045] Figure 1 This is a schematic diagram of an application scenario of the attribute recognition pre-trained model generation method according to some embodiments of the present disclosure.

[0046] exist Figure 1In this application scenario, the electronic device 101 first acquires first item information for the first target item 102. This first item information includes: a first item image 103 and multiple pre-stored item attribute information 104. In this application scenario, the multiple pre-stored item attribute information 104 includes: pre-stored item attribute information 1041, pre-stored item attribute information 1042, pre-stored item attribute information 1043, and pre-stored item attribute information 1044. Pre-stored item attribute information 1041 can be "Color: Green". Pre-stored item attribute information 1042 can be "Category: Green Tea". Pre-stored item attribute information 1043 can be "Price: 123 yuan". Pre-stored item attribute information 1044 can be "Weight: 500g". Then, the electronic device 101 can input the first item image 103 into the image feature extraction model 106 included in the initial attribute recognition model 105 to obtain a first image feature matrix 107, and input each of the pre-stored item attribute information 104 from the multiple pre-stored item attribute information to the attribute feature extraction model 108 included in the initial attribute recognition model 105 to generate a first attribute feature vector, resulting in multiple first attribute feature vectors 109. In this application scenario, the first image feature matrix 107 can be "[[1,2,3,42,123],[3,4,2,12,533],[7,12,34,2,3]]". The multiple first attribute feature vectors 109 can include: the first attribute feature vector 1091 corresponding to the pre-stored item attribute information 1041, the first attribute feature vector 1092 corresponding to the pre-stored item attribute information 1042, the first attribute feature vector 1093 corresponding to the pre-stored item attribute information 1043, and the first attribute feature vector 1094 corresponding to the pre-stored item attribute information 1044. The first attribute feature vector 1091 can be “[4,2,13,53,66]”. The second attribute feature vector 1092 can be “[9,3,42,123,3]”. The third attribute feature vector 1093 can be “[89,34,2,3,12]”. The fourth attribute feature vector 1094 can be “[55,33,12,4,3]”. Next, the electronic device 101 can input the aforementioned multiple first attribute feature vectors 109 into the masking processing model 110 included in the initial attribute recognition model 105 to perform masking processing on at least one of the multiple first attribute feature vectors 109, obtaining multiple masked attribute feature vectors 111. In this application scenario, the multiple masked attribute feature vectors 111 can include: masked attribute feature vector 1111, masked attribute feature vector 1112, masked attribute feature vector 1113, and masked attribute feature vector 1114. The masked attribute feature vector 1111 can be "[4,2,13,53,66]". The masked attribute feature vector 1112 can be "[M,M,M,M,M]".The masked attribute feature vector 1113 can be "[89,34,2,3,12]". The masked attribute feature vector 1114 can be "[M,M,M,M,M]". At least one first attribute feature vector can include: first attribute feature vector 1092 and first attribute feature vector 1094. Furthermore, the electronic device 101 can input the aforementioned first image feature matrix 107 and the aforementioned plurality of masked attribute feature vectors 111 into the masked attribute information generation model 112 included in the aforementioned initial attribute recognition model 105 to generate at least one masked attribute information 113. In this application scenario, at least one masked attribute information 113 can include: masked attribute information 1131 corresponding to masked attribute feature vector 1112 and masked attribute information 1132 corresponding to masked attribute feature vector 1114. Masked attribute information 1131 can be "Category: Green Tea". Masked attribute information 1132 can be "Weight: 510g". Finally, the electronic device 101 can train the initial attribute recognition model 105 based on the at least one first attribute feature vector and the at least one mask attribute information 113 to obtain the trained attribute recognition model, which serves as the attribute recognition pre-training model.

[0047] It should be noted that the aforementioned electronic device 101 can be either hardware or software. When the electronic device is hardware, it can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or as a single server or a single terminal device. When the electronic device is software, it can be installed in the hardware devices listed above. It can be implemented as, for example, multiple software programs or software modules used to provide distributed services, or as a single software program or software module. No specific limitations are made here.

[0048] It should be understood that Figure 1 The number of electronic devices shown is merely illustrative. Any number of electronic devices can be used depending on the implementation requirements.

[0049] Continue to refer to Figure 2 The flowchart 200 illustrates some embodiments of the attribute recognition pre-trained model generation method according to the present disclosure. The attribute recognition pre-trained model generation method includes the following steps:

[0050] Step 201: Obtain the first item information for the first target item.

[0051] In some embodiments, the execution entity of the above-described attribute recognition pre-trained model generation method (e.g.) Figure 1The electronic device 101 shown can acquire first item information for a first target item via a wired or wireless connection. The first item information includes: a first item image and multiple pre-stored item attribute information. The first item image can be an image of the first item. The pre-stored item attribute information can be pre-stored item attribute information specific to the first target item. The multiple pre-stored item attribute information can be all item attribute information, or it may not be all item attribute information. All item attribute information can be attribute information for all item attributes. Item attribute information can be attribute values ​​of item attributes. For example, item attributes can be, but are not limited to, at least one of the following: item model, item color, item size, and item value.

[0052] For example, in an e-commerce scenario, the primary target item can be a product. The primary item image can be a product image. Multiple pre-stored item attribute information can include: product model information, product color information, product size information, and product price information.

[0053] Step 202: Input the first item image into the image feature extraction model included in the initial attribute recognition model to obtain the first image feature matrix, and input each of the multiple pre-stored item attribute information into the attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining multiple first attribute feature vectors.

[0054] In some embodiments, the executing entity may input the first item image into the image feature extraction model included in the initial attribute recognition model to obtain a first image feature matrix, and input each of the multiple pre-stored item attribute information into the attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining multiple first attribute feature vectors. The initial item attribute recognition model may be an item attribute recognition model whose model parameters have not been fully updated. The item attribute recognition model may be a neural network model for recognizing item attribute information. For example, the item attribute recognition model may be a multi-parallel-connected convolutional neural network. The image feature extraction model may be a model for extracting image feature information. For example, the image feature extraction model may be a convolutional neural network. The attribute feature extraction model may be a model for extracting attribute feature information. For example, the attribute feature extraction model may be a convolutional neural network.

[0055] Step 203: Input the above-mentioned multiple first attribute feature vectors into the masking processing model included in the above-mentioned initial attribute recognition model, so as to perform masking processing on at least one of the above-mentioned multiple first attribute feature vectors to obtain multiple masked attribute feature vectors.

[0056] In some embodiments, the execution entity may input the plurality of first attribute feature vectors into a masking processing model included in the initial attribute recognition model to perform masking processing on at least one of the plurality of first attribute feature vectors, obtaining a plurality of masked attribute feature vectors. The masking processing model may be a model for masking attribute feature vectors. For example, the first attribute feature vectors may be "[3, 4, 6, 7, 13]". The masked attribute feature vectors may be "[M, M, M, M, M]". At least one first attribute feature vector has at least one corresponding masked attribute feature vector. For example, the masking processing model may be a convolutional neural network.

[0057] It should be noted that the number of first attribute feature vectors included in the above-mentioned multiple first attribute feature vector sets may be greater than or equal to the number of attribute feature vectors included in at least one attribute feature vector set.

[0058] As an example, firstly, the aforementioned execution entity can use a masking model to randomly select at least one attribute feature vector from the multiple first attribute feature vectors. Then, the execution entity can use the masking model to mask the at least one attribute feature vector, obtaining at least one masked attribute feature vector. Finally, the subset of first attribute feature vectors that were not selected from the multiple first attribute feature vectors and the at least one masked attribute feature vector are fused to obtain a fused vector set, which serves as the multiple masked attribute feature vectors.

[0059] Step 204: Input the first image feature matrix and the multiple masked attribute feature vectors into the masked attribute information generation model included in the initial attribute recognition model to generate at least one masked attribute information.

[0060] In some embodiments, the execution entity may input the first image feature matrix and the plurality of masked attribute feature vectors into the masked attribute information generation model included in the initial attribute recognition model to generate at least one masked attribute information. The at least one masked attribute information may be attribute prediction information for the at least one masked attribute feature vector. The masked attribute information generation model may be a model that generates corresponding attribute prediction information for the masked attribute feature vectors. For example, the masked attribute information generation model may be a Transformer model or a Graph Convolutional Network (GCN) model.

[0061] Step 205: Based on the above-mentioned at least one first attribute feature vector and the above-mentioned at least one mask attribute information, train the above-mentioned initial attribute recognition model to obtain the trained attribute recognition model, which serves as the attribute recognition pre-training model.

[0062] In some embodiments, the execution entity may train the initial attribute recognition model based on at least one first attribute feature vector and at least one mask attribute information to obtain a trained attribute recognition model, which serves as a pre-trained attribute recognition model.

[0063] As an example, firstly, the execution entity can perform vector transformation on each of the at least one mask attribute information to generate a mask vector, thus obtaining at least one mask vector. Then, the execution entity can determine the similarity between the at least one mask vector and corresponding vectors in the at least one first attribute feature vector, thus obtaining at least one similarity. Next, the execution entity can determine the variance corresponding to the at least one similarity. Finally, the variance is used as the loss value to train the initial attribute recognition model, resulting in a trained attribute recognition model, which serves as the pre-trained attribute recognition model.

[0064] In some optional implementations of certain embodiments, training the initial attribute recognition model based on at least one first attribute feature vector and at least one mask attribute information to obtain a trained attribute recognition model, which serves as a pre-trained attribute recognition model, may include the following steps:

[0065] The first step is to obtain at least one first attribute information corresponding to at least one first attribute feature vector.

[0066] There is a one-to-one correspondence between the first attribute feature vector in at least one first attribute feature vector and the first attribute information in at least one first attribute information.

[0067] The second step is to generate a loss value for the first target item based on at least one first attribute information and at least one mask attribute information.

[0068] As an example, firstly, the executing entity can determine at least one attribute vector corresponding to at least one first attribute information. Then, the executing entity can determine at least one mask attribute vector corresponding to at least one mask attribute information. Finally, the executing entity can input the at least one attribute vector and the at least one mask attribute vector into a cross-entropy loss function to generate a loss value for the first target item.

[0069] The third step is to train the initial attribute recognition model based on the loss value to obtain the trained attribute recognition model, which serves as the pre-trained attribute recognition model.

[0070] As an example, in response to determining that the loss value is less than or equal to a preset value, the initial attribute recognition model is determined as the trained attribute recognition model, serving as the attribute recognition pre-trained model. In response to determining that the loss value is greater than the preset value, the model parameters in the initial attribute recognition model can be updated based on the loss value to obtain an updated model, which serves as the attribute recognition pre-trained model.

[0071] The above embodiments of this disclosure have the following beneficial effects: the attribute recognition pre-training model generation method of some embodiments of this disclosure can generate an attribute recognition pre-training model with more accurate attribute information. Specifically, the reason why the attribute recognition pre-training model is not accurate enough in identifying attribute information is that the attribute recognition model only uses the image feature information of the item to identify the item attribute information, which has the problems of insufficient recognition accuracy and poor recognition effect. Based on this, the attribute recognition pre-training model generation method of some embodiments of this disclosure firstly obtains first item information for a first target item. The first item information includes: a first item image and multiple pre-stored item attribute information. Here, the first item information is obtained as the data basis for subsequent model training. In addition, for the attribute information annotation of the first target item, the problem of a large amount of data to be annotated can be greatly reduced by using multiple pre-stored item attribute information. Next, the first item image is input into the image feature extraction model included in the initial attribute recognition model to obtain a first image feature matrix, and each of the multiple pre-stored item attribute information is input into the attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining multiple first attribute feature vectors. Here, image feature information and attribute feature information are extracted for subsequent attribute recognition models to learn the correspondence between image features and attribute features, and the correspondence between attribute features themselves. Next, the aforementioned multiple first attribute feature vectors are input into the masking model included in the initial attribute recognition model to mask at least one of the multiple first attribute feature vectors, obtaining multiple masked attribute feature vectors. Masking at least one first attribute feature vector facilitates the subsequent determination of the correspondence between attribute features. Then, the aforementioned first image feature matrix and the aforementioned multiple masked attribute feature vectors are input into the masked attribute information generation model included in the initial attribute recognition model to generate at least one masked attribute information. Here, the masked attribute information generation model can effectively learn not only the association between the first image feature matrix and each masked attribute feature vector, but also the attribute association between the various masked attribute feature vectors. Finally, based on the aforementioned at least one first attribute feature vector and the aforementioned at least one masked attribute information, the initial attribute recognition model is trained to obtain a trained attribute recognition model, which serves as a pre-trained model for attribute recognition. The resulting attribute recognition pre-trained model can effectively and accurately identify item attribute information by considering the correspondence between image features and attribute features, as well as the correspondence between attribute features themselves. In summary, the attribute recognition model, during training, not only considers the correspondence between image features and attribute features, but also fully considers the correspondence between attribute features themselves.Therefore, a more accurate attribute recognition pre-trained model can be obtained.

[0072] Further reference Figure 3 The diagram illustrates flow 300 of some other embodiments of the attribute recognition pre-trained model generation method according to the present disclosure. This attribute recognition pre-trained model generation method includes the following steps:

[0073] Step 301: Obtain the first item information for the first target item.

[0074] Step 302: Input the first item image into the image feature extraction model included in the initial attribute recognition model to obtain the first image feature matrix, and input each of the multiple pre-stored item attribute information into the attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining multiple first attribute feature vectors.

[0075] Step 303: Input the above-mentioned multiple first attribute feature vectors into the masking processing model included in the above-mentioned initial attribute recognition model, so as to perform masking processing on at least one of the above-mentioned multiple first attribute feature vectors to obtain multiple masked attribute feature vectors.

[0076] In some embodiments, the specific implementation of steps 301-303 and the resulting technical effects can be found in [reference needed]. Figure 2 Steps 201-203 in the corresponding embodiments will not be repeated here.

[0077] Step 304: Perform a linear transformation on the first image feature matrix to generate a transformation vector set.

[0078] In some embodiments, the executing entity (e.g. Figure 1 The electronic device 101 shown can linearly transform the first image feature matrix to generate a set of transformation vectors.

[0079] As an example, firstly, the aforementioned execution entity can determine the number of attributes corresponding to the first target item. Then, based on the number of attributes, the aforementioned first image feature matrix is ​​linearly transformed into multiple vectors to generate a transformation vector set.

[0080] Step 305: Perform vector fusion between the modified vector and the multiple masked attribute feature vectors to obtain the first fused vector.

[0081] In some embodiments, the execution entity may perform vector fusion with the multiple masked attribute feature vectors to obtain a first fused vector.

[0082] Step 306: Input the first fusion vector into the mask vector generation model included in the mask attribute information generation model to generate at least one mask prediction vector.

[0083] In some embodiments, the execution entity may input the first fusion vector into a mask vector generation model included in the mask attribute information generation model to generate at least one mask prediction vector. The mask vector generation model may be a model for generating mask prediction vectors. The mask prediction vector may be a predicted attribute feature vector corresponding to the masked attribute feature vector. For example, the first attribute feature vector may be "[3, 4, 6, 7, 13]". The masked attribute feature vector may be "[M, M, M, M, M]". Then the mask prediction vector may be "[3, 4, 5, 7, 13]". For example, the mask vector generation model may be a convolutional neural network.

[0084] Step 307: Input each of the at least one mask prediction vectors into the classification model included in the mask attribute information generation model to generate mask attribute information, thereby obtaining the at least one mask attribute information.

[0085] In some embodiments, the execution entity may input each of the at least one mask prediction vector into a classification model included in the mask attribute information generation model to generate mask attribute information, thereby obtaining at least one mask attribute information. The classification model may be a model used to determine the attribute information corresponding to the mask prediction vector. For example, the classification model may be a convolutional neural network.

[0086] Step 308: Based on the above-mentioned at least one first attribute feature vector and the above-mentioned at least one mask attribute information, train the above-mentioned initial attribute recognition model to obtain the trained attribute recognition model, which serves as the attribute recognition pre-training model.

[0087] In some embodiments, the specific implementation of step 308 and its resulting technical effects can be found in [reference needed]. Figure 2 Step 204 in the corresponding embodiment will not be repeated here.

[0088] from Figure 3 It can be seen from this that, with Figure 2 Compared to the description of some corresponding embodiments, Figure 3The flowchart 300 of the attribute recognition pre-training model generation method in some corresponding embodiments discloses the specific steps of the mask attribute information generation model to generate at least one mask attribute information. Here, by utilizing the mask vector generation model in the mask attribute information generation model, the mask prediction vector can be accurately generated effectively based on the correspondence between image features and attribute features, and the correspondence between attribute features. Based on the generated at least one mask prediction vector, a classification model can generate at least one more accurate mask attribute information.

[0089] Continue to refer to Figure 4 The diagram illustrates a flow 400 of some embodiments of the attribute recognition model generation method according to the present disclosure. The attribute recognition model generation method includes the following steps:

[0090] Step 401: Obtain the attribute recognition pre-trained model and the item labeling information for the second target item.

[0091] In some embodiments, the execution entity of the above-described attribute recognition model generation method (e.g.) Figure 1 The electronic device 101 shown can acquire an attribute recognition pre-trained model and item annotation information for the second target item. The attribute recognition pre-trained model can be an attribute recognition model pre-trained upstream based on pre-stored item attribute information. The item annotation information can be annotation information for labeling the attributes of the second target item.

[0092] Step 402: Determine the task information corresponding to the downstream task.

[0093] In some embodiments, the executing entity may determine the task information corresponding to the downstream task. The downstream task is a task associated with item attributes. That is, the downstream task can be a task involving item attributes. For example, the downstream task can be an image search task or an item detection task. The task information may be the task content information of the downstream task or the task identifier of the downstream task.

[0094] Step 403: In response to determining that the above task information is attribute recognition task information, the above attribute recognition pre-trained model is subjected to a first model structure adjustment to generate an adjusted attribute recognition model.

[0095] In some embodiments, in response to determining that the task information is attribute recognition task information, the executing entity may perform a first model structure adjustment on the attribute recognition pre-trained model to generate an adjusted attribute recognition model. The attribute recognition task information may be task content information of the attribute recognition task, or it may be identification information of the attribute recognition task.

[0096] As an example, the aforementioned execution entity can add a predetermined number of image feature extraction models after the image feature extraction model to extract image feature information in greater depth. Similarly, the aforementioned execution entity can add a predetermined number of attribute feature extraction models after the attribute feature extraction model to extract attribute feature information in greater depth.

[0097] In some optional implementations of certain embodiments, the execution entity may remove the mask processing model included in the attribute recognition pre-trained model and modify the mask attribute information generation model included in the attribute recognition pre-trained model to obtain an adjusted attribute recognition model.

[0098] As an example, the executing entity can modify the mask attribute information generation model into a vector fusion model, a related feature information extraction model, and a classification model. The vector fusion model can be a model that performs vector fusion. The related feature information extraction model can be a model that extracts the relationships between image features and attribute features, and the relationships between attribute features themselves. For example, the related feature information extraction model can be a Transformer model. The output of the related feature information extraction model can be in vector form. The classification model can be a model that generates attribute information based on the related feature vectors. The related feature vectors can be the output of the related feature information extraction model.

[0099] Step 404: Based on the above item labeling information, train the adjusted attribute recognition model to obtain the attribute recognition model.

[0100] In some embodiments, the aforementioned execution entity may train the aforementioned adjusted attribute recognition model based on the aforementioned item labeling information to obtain the attribute recognition model.

[0101] In some optional implementations of certain embodiments, after step 404, the steps further include:

[0102] The first step, in response to the determination that the aforementioned task information is not attribute recognition task information, involves a second model structure adjustment of the attribute recognition pre-trained model based on the task information, to generate an initial task model for the aforementioned task information. The initial task model can be a model whose model parameters have not yet been fully updated. The aforementioned task model can be a model that performs the task corresponding to the task information. The task information is the task content information of a graph search task. The aforementioned task model can be a graph search task model.

[0103] The second step is to train the initial task model based on the item labeling information to obtain the trained task model.

[0104] As an example, the aforementioned execution entity can use the item labeling information as training data to train the initial task model and obtain the trained task model.

[0105] In some optional implementations of certain embodiments, the aforementioned item annotation information includes: a second item image of the second target item and multiple item annotation attribute information. The item annotation attribute information may be annotation information for annotating the attributes of the second target item.

[0106] Optionally, training the adjusted attribute recognition model based on the aforementioned item labeling information to obtain the attribute recognition model may include the following steps:

[0107] The first step involves inputting the image of the second item into the image feature extraction model included in the initially adjusted attribute recognition model to obtain the second image feature matrix, and inputting each item label attribute information from the multiple item label attribute information into the attribute feature extraction model included in the adjusted attribute recognition model to generate a second attribute feature vector, thus obtaining multiple second attribute feature vectors.

[0108] The image feature extraction model described above can be a model for extracting image feature information. For example, an image feature extraction model can be a convolutional neural network. The attribute feature extraction model can be a model for extracting attribute feature information. For example, an attribute feature extraction model can be a convolutional neural network.

[0109] The second step is to perform vector fusion of the second image feature matrix and the multiple second attribute feature vectors to generate a second fusion vector.

[0110] The third step is to input the second fusion vector into the attribute association information generation model included in the initially adjusted attribute recognition model to generate attribute association feature vectors.

[0111] The attribute association information generation model can be a model that generates information about the relationships between various attributes. For example, the attribute association information generation model mentioned above could be a Transformer model.

[0112] The fourth step is to input the aforementioned attribute-associated feature vectors into the classification model included in the initially adjusted attribute recognition model to generate multiple predicted attribute information.

[0113] The classification model can be an attribute information classification model. For example, the classification model mentioned above could be a convolutional neural network.

[0114] The fifth step is to obtain multiple item attribute information for the second target item mentioned above.

[0115] Among them, there is a one-to-one correspondence between the true attribute information of multiple items and the predicted attribute information of multiple predicted attribute information.

[0116] The sixth step is to train the adjusted attribute recognition model based on the predicted attribute information and the actual attribute information of the items mentioned above, and obtain the attribute recognition model.

[0117] As an example, firstly, the executing entity can determine multiple predicted attribute vectors corresponding to the multiple predicted attribute information. Then, the executing entity can determine multiple true attribute vectors corresponding to the multiple true attribute information of the items. Next, the executing entity can input the multiple predicted attribute vectors and the multiple true attribute vectors of the items into the cross-entropy loss function to generate a loss value for the second target item. Finally, based on the loss value, the adjusted attribute recognition model is trained to obtain the attribute recognition model.

[0118] The above embodiments of this disclosure have the following beneficial effects: The attribute recognition model generation method of some embodiments of this disclosure first obtains an attribute recognition pre-trained model and item annotation information for a second target item. Obtaining the attribute recognition pre-trained model facilitates adjustments to its model structure for downstream tasks. Obtaining the item annotation information for the second target item facilitates subsequent model training. Next, the task information corresponding to the downstream task is determined. The downstream task is a task related to item attributes. Determining the task information of the downstream task facilitates determining how to adjust the model structure of the attribute recognition pre-trained model to generate a task model for the task information. Furthermore, in response to determining that the task information is attribute recognition task information, the attribute recognition pre-trained model undergoes a first model structure adjustment to generate an adjusted attribute recognition model, facilitating subsequent execution of downstream tasks. Finally, based on the item annotation information, the adjusted attribute recognition model is trained to obtain an attribute recognition model that generates more accurate attribute recognition information.

[0119] Continue to refer to Figure 5 The flowchart 500 illustrates some embodiments of the full item attribute information generation method according to this disclosure. The full item attribute information generation method includes the following steps:

[0120] Step 501: Obtain information about the third item for the third target item.

[0121] In some embodiments, the execution subject of the above-described full item attribute information generation method (e.g.) Figure 1The electronic device 101 shown can acquire third item information for a third target item via wired or wireless means. The third item information includes: a third item image and multiple pre-stored item attribute information. The third item image can be an image of the third item. The pre-stored item attribute information can be pre-stored item attribute information specific to the third target item. The multiple pre-stored item attribute information can be all item attribute information, or it can be less than all item attribute information. All item attribute information can be attribute information for all item attributes of the item.

[0122] Step 502: Input the aforementioned third item information into the pre-trained attribute recognition model to output the full item attribute information for the aforementioned third target item.

[0123] In some embodiments, the executing entity may input the third item information into a pre-trained attribute recognition model to output full item attribute information for the third target item. The attribute recognition model is generated using the attribute recognition pre-training model generation method and attribute recognition model generation method of some embodiments of this disclosure.

[0124] The above embodiments of this disclosure have the following beneficial effects: the full item attribute information generation method of some embodiments of this disclosure can accurately generate more comprehensive full item attribute information for a third target item.

[0125] Further reference Figure 6 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of an attribute recognition pre-trained model generation device, which are similar to... Figure 2 Corresponding to the method embodiments shown, this attribute recognition pre-trained model generation device can be specifically applied to various electronic devices.

[0126] like Figure 6As shown, an attribute recognition pre-trained model generation device 600 includes: a first acquisition unit 601, a first input unit 602, a second input unit 603, a third input unit 604, and a first training unit 605. The first acquisition unit 601 is configured to acquire first item information for a first target item, wherein the first item information includes: a first item image and multiple pre-stored item attribute information; the first input unit 602 is configured to input the first item image into an image feature extraction model included in the initial attribute recognition model to obtain a first image feature matrix, and to input each of the multiple pre-stored item attribute information into the attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining multiple first attribute feature vectors; the second input unit 603 is configured to input the multiple first attribute feature vectors into a... The initial attribute recognition model includes a masking processing model to mask at least one of the plurality of first attribute feature vectors to obtain a plurality of masked attribute feature vectors; the third input unit 604 is configured to input the first image feature matrix and the plurality of masked attribute feature vectors into the mask attribute information generation model included in the initial attribute recognition model to generate at least one mask attribute information; the first training unit 605 is configured to train the initial attribute recognition model based on the at least one first attribute feature vector and the at least one mask attribute information to obtain a trained attribute recognition model, which serves as a pre-trained attribute recognition model.

[0127] In some optional implementations of certain embodiments, the third input unit 604 in the attribute recognition pre-trained model generation device 600 may be further configured to: linearly transform the first image feature matrix to generate a transformation vector set; fuse the transformation vector set with the plurality of masked attribute feature vectors to obtain a first fused vector; input the first fused vector to a mask vector generation model included in the mask attribute information generation model to generate at least one mask prediction vector; and input each of the at least one mask prediction vectors to a classification model included in the mask attribute information generation model to generate mask attribute information, thereby obtaining at least one mask attribute information.

[0128] In some optional implementations of some embodiments, the first training unit 605 in the attribute recognition pre-training model generation device 600 may be further configured to: acquire at least one first attribute information corresponding to the at least one first attribute feature vector; generate a loss value for the first target item based on the at least one first attribute information and the at least one mask attribute information; and train the initial attribute recognition model based on the loss value to obtain a trained attribute recognition model as an attribute recognition pre-training model.

[0129] It is understandable that the units described in the attribute recognition pre-trained model generation device 600 and the reference Figure 2 The steps in the described method correspond accordingly. Therefore, the operations, features, and beneficial effects described above for the method also apply to the attribute recognition pre-trained model generation device 600 and the units contained therein, and will not be repeated here.

[0130] Further reference Figure 7 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of an attribute recognition model generation apparatus, which are similar to... Figure 4 Corresponding to the method embodiments shown, this attribute recognition model generation device can be specifically applied to various electronic devices.

[0131] like Figure 7 As shown, an attribute recognition model generation device 700 includes: a second acquisition unit 701, a determination unit 702, a model structure adjustment unit 703, and a second training unit 704. The second acquisition unit 701 is configured to acquire an attribute recognition pre-trained model and item labeling information for a second target item; the determination unit 702 is configured to determine task information corresponding to a downstream task, wherein the downstream task is a task associated with item attributes; the model structure adjustment unit 703 is configured to, in response to determining that the task information is attribute recognition task information, perform a first model structure adjustment on the attribute recognition pre-trained model to generate an adjusted attribute recognition model; the second training unit 704 is configured to train the adjusted attribute recognition model based on the item labeling information to obtain the attribute recognition model.

[0132] In some optional implementations of certain embodiments, the attribute recognition model generation apparatus 700 may further include an adjustment unit and a third training unit (not shown in the figure). The adjustment unit may be configured to: in response to determining that the task information is not attribute recognition task information, adjust the model structure of the attribute recognition pre-trained model according to the task information to generate an initial task model for the task information. The third training unit may be configured to: train the initial task model according to the item labeling information to obtain a trained task model.

[0133] In some optional implementations of some embodiments, the model structure adjustment unit 703 in the attribute recognition model generation device 700 may be further configured to: remove the mask processing model included in the attribute recognition pre-trained model, and modify the mask attribute information generation model included in the attribute recognition pre-trained model to obtain the adjusted attribute recognition model.

[0134] In some optional implementations of certain embodiments, the aforementioned item labeling information includes: a second item image of the second target item and multiple item labeling attribute information; and the second training unit 704 in the attribute recognition model generation device 700 can be further configured to: input the aforementioned second item image into the image feature extraction model included in the initially adjusted attribute recognition model to obtain a second image feature matrix, and input each item labeling attribute information in the multiple item labeling attribute information into the attribute feature extraction model included in the adjusted attribute recognition model to generate a second attribute feature vector, thereby obtaining multiple second attribute feature vectors; and process the aforementioned second image feature matrix. A linear transformation is performed to obtain a second transformation vector set; the second transformation vector set and the multiple second attribute feature vectors are fused to generate a second fused vector; the second fused vector is input into the attribute association information generation model included in the initially adjusted attribute recognition model to generate attribute association feature vectors; the attribute association feature vectors are input into the classification model included in the initially adjusted attribute recognition model to generate multiple predicted attribute information; multiple real attribute information of the second target item is obtained; based on the multiple predicted attribute information and the multiple real attribute information of the item, the adjusted attribute recognition model is trained to obtain the attribute recognition model.

[0135] It is understandable that the units described in the attribute recognition model generation device 700 and the reference Figure 4 The steps in the described method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the attribute recognition model generation device 700 and the units contained therein, and will not be repeated here.

[0136] Further reference Figure 8 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a device for generating full-volume item attribute information. These device embodiments are similar to... Figure 5 Corresponding to the method embodiments shown, this full-item attribute information generation device can be specifically applied to various electronic devices.

[0137] like Figure 8 As shown, a full-item attribute information generation device 800 includes a third acquisition unit 801 and a fourth input unit 802. The third acquisition unit 801 is configured to acquire third item information for a third target item; the fourth input unit 802 is configured to input the third item information into a pre-trained attribute recognition model to output full-item attribute information for the third target item. The attribute recognition model is generated using the attribute recognition pre-training model generation method and the attribute recognition model generation method according to some embodiments of this disclosure.

[0138] It is understandable that the units recorded in the full-item attribute information generation device 800 are related to the reference. Figure 5 The steps in the described method correspond accordingly. Therefore, the operations, features, and beneficial effects described above for the method also apply to the full-item attribute information generation device 800 and the units contained therein, and will not be repeated here.

[0139] The following is for reference. Figure 9 It illustrates electronic devices suitable for implementing some embodiments of this disclosure (e.g., Figure 1 A schematic diagram of the structure of electronic device 101)900 in the middle. Figure 9 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0140] like Figure 9 As shown, electronic device 900 may include a processing device (e.g., a central processing unit, a graphics processor, etc.) 901, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 902 or a program loaded from storage device 908 into random access memory (RAM) 903. RAM 903 also stores various programs and data required for the operation of electronic device 900. Processing device 901, ROM 902, and RAM 903 are interconnected via bus 904. Input / output (I / O) interface 905 is also connected to bus 904.

[0141] Typically, the following devices can be connected to I / O interface 905: input devices 906 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 907 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 908 including, for example, magnetic tapes, hard disks, etc.; and communication devices 909. Communication device 909 allows electronic device 900 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 9 An electronic device 900 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 9 Each box shown can represent a device or multiple devices as needed.

[0142] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 909, or installed from a storage device 908, or installed from a ROM 902. When the computer program is executed by the processing device 901, it performs the functions defined in the methods of some embodiments of this disclosure.

[0143] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0144] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol, such as HTTP (Hypertext Transfer Protocol), and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0145] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: acquire first item information for a first target item, wherein the first item information includes: a first item image and multiple pre-stored item attribute information; input the first item image into an image feature extraction model included in an initial attribute recognition model to obtain a first image feature matrix; and input each of the multiple pre-stored item attribute information into an attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining multiple first attribute feature vectors; The aforementioned multiple first attribute feature vectors are input into the masking processing model included in the initial attribute recognition model to perform masking processing on at least one of the multiple first attribute feature vectors, resulting in multiple masked attribute feature vectors. The aforementioned first image feature matrix and the aforementioned multiple masked attribute feature vectors are input into the masked attribute information generation model included in the initial attribute recognition model to generate at least one masked attribute information. Based on the aforementioned at least one first attribute feature vector and the aforementioned at least one masked attribute information, the initial attribute recognition model is trained to obtain a trained attribute recognition model, which serves as the attribute recognition pre-training model. The attribute recognition pre-training model and the item labeling information for the second target item are obtained. The task information corresponding to the downstream task is determined, wherein the downstream task is a task associated with the item attribute. In response to determining that the task information is attribute recognition task information, the attribute recognition pre-training model undergoes a first model structure adjustment to generate an adjusted attribute recognition model. Based on the aforementioned item labeling information, the adjusted attribute recognition model is trained to obtain the attribute recognition model. Obtain third item information for a third target item; input the third item information into a pre-trained attribute recognition model to output full item attribute information for the third target item, wherein the attribute recognition model is generated by the attribute recognition pre-trained model generation method and the attribute recognition model generation method of some embodiments of this disclosure.

[0146] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0147] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0148] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including a first acquisition unit, a first input unit, a second input unit, a third input unit, and a first training unit. The names of these units do not necessarily limit the specific unit; for example, the first acquisition unit may also be described as "a unit for acquiring first item information for a first target item."

[0149] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0150] Some embodiments of this disclosure also provide a computer program product, including a computer program that, when executed by a processor, implements any of the above-described methods for generating attribute recognition pre-trained models, attribute recognition models, and full-item attribute information.

[0151] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for generating a pre-trained model for attribute recognition, comprising: Obtain first item information for a first target item, wherein the first item information includes: a first item image and multiple pre-stored item attribute information; The first item image is input into the image feature extraction model included in the initial attribute recognition model to obtain the first image feature matrix, and each of the multiple pre-stored item attribute information is input into the attribute feature extraction model included in the initial attribute recognition model to generate the first attribute feature vector, thereby obtaining multiple first attribute feature vectors. The plurality of first attribute feature vectors are input into the masking processing model included in the initial attribute recognition model to perform masking processing on at least one of the plurality of first attribute feature vectors to obtain a plurality of masked attribute feature vectors. The first image feature matrix and the plurality of masked attribute feature vectors are input into the mask attribute information generation model included in the initial attribute recognition model to generate at least one mask attribute information. The initial attribute recognition model is trained based on the at least one first attribute feature vector and the at least one mask attribute information to obtain a trained attribute recognition model, which serves as the attribute recognition pre-training model.

2. The method according to claim 1, wherein, The step of inputting the first image feature matrix and the plurality of masked attribute feature vectors into the masked attribute information generation model included in the initial attribute recognition model to generate at least one masked attribute information includes: The first image feature matrix is ​​linearly transformed to generate a set of transformed vectors; The change vector set is fused with the multiple masked attribute feature vectors to obtain a first fused vector. The first fusion vector is input into the mask vector generation model included in the mask attribute information generation model to generate the at least one mask prediction vector. Each of the at least one mask prediction vectors is input into the classification model included in the mask attribute information generation model to generate mask attribute information, thereby obtaining the at least one mask attribute information.

3. The method according to claim 1, wherein, The step of training the initial attribute recognition model based on the at least one first attribute feature vector and the at least one mask attribute information to obtain a trained attribute recognition model, which serves as a pre-trained model for attribute recognition, includes: Obtain at least one first attribute information corresponding to the at least one first attribute feature vector; Based on the at least one first attribute information and the at least one mask attribute information, a loss value for the first target item is generated; Based on the loss value, the initial attribute recognition model is trained to obtain the trained attribute recognition model, which serves as the pre-trained attribute recognition model.

4. A method for generating an attribute recognition model, comprising: Obtain an attribute recognition pre-trained model and item labeling information for a second target item, wherein the attribute recognition pre-trained model is generated based on the method described in any one of claims 1-3; Determine the task information corresponding to the downstream task, wherein the downstream task is a task associated with the item attribute; In response to determining that the task information is attribute recognition task information, the attribute recognition pre-trained model is subjected to a first model structure adjustment to generate an adjusted attribute recognition model; Based on the item labeling information, the adjusted attribute recognition model is trained to obtain the attribute recognition model.

5. The method according to claim 4, wherein, The method further includes: In response to determining that the task information is not attribute recognition task information, the model structure of the attribute recognition pre-trained model is adjusted in a second way according to the task information to generate an initial task model for the task information. The initial task model is trained based on the item labeling information to obtain the trained task model.

6. The method according to claim 4, wherein, The step of adjusting the first model structure of the attribute recognition pre-trained model to generate the adjusted attribute recognition model includes: The masking processing model included in the attribute recognition pre-trained model is removed, and the mask attribute information generation model included in the attribute recognition pre-trained model is modified to obtain the adjusted attribute recognition model.

7. The method according to claim 6, wherein, The item labeling information includes: a second item image of the second target item and multiple item labeling attribute information; and The step of training the adjusted attribute recognition model based on the item labeling information to obtain the attribute recognition model includes: The second item image is input into the image feature extraction model included in the adjusted attribute recognition model to obtain a second image feature matrix, and each item label attribute information in the multiple item label attribute information is input into the attribute feature extraction model included in the adjusted attribute recognition model to generate a second attribute feature vector, thereby obtaining multiple second attribute feature vectors; A second image feature matrix is ​​linearly transformed to obtain a second transformation vector set; The second set of change vectors and the plurality of second attribute feature vectors are fused to generate a second fused vector; The second fusion vector is input into the attribute association information generation model included in the adjusted attribute recognition model to generate an attribute association feature vector; The attribute-associated feature vector is input into the classification model included in the adjusted attribute recognition model to generate multiple predicted attribute information. Obtain multiple real attribute information of the second target item; Based on the multiple predicted attribute information and the multiple actual attribute information of the items, the adjusted attribute recognition model is trained to obtain the attribute recognition model.

8. A method for generating full item attribute information, comprising: Obtain information about the third item in relation to the third target item; The third item information is input into a pre-trained attribute recognition model to output full item attribute information for the third target item, wherein the attribute recognition model is generated based on the method described in any one of claims 4-7.

9. An attribute recognition pre-trained model generation device, comprising: The first acquisition unit is configured to acquire first item information for a first target item, wherein the first item information includes: a first item image and multiple pre-stored item attribute information; The first input unit is configured to input the first item image into the image feature extraction model included in the initial attribute recognition model to obtain a first image feature matrix, and to input each of the multiple pre-stored item attribute information into the attribute feature extraction model included in the initial attribute recognition model to generate a first attribute feature vector, thereby obtaining multiple first attribute feature vectors. The second input unit is configured to input the plurality of first attribute feature vectors into the masking processing model included in the initial attribute recognition model, so as to perform masking processing on at least one of the plurality of first attribute feature vectors to obtain a plurality of masked attribute feature vectors. The third input unit is configured to input the first image feature matrix and the plurality of masked attribute feature vectors into the mask attribute information generation model included in the initial attribute recognition model to generate at least one mask attribute information. The first training unit is configured to train the initial attribute recognition model based on the at least one first attribute feature vector and the at least one mask attribute information to obtain a trained attribute recognition model, which serves as the attribute recognition pre-training model.

10. An attribute recognition model generation device, comprising: The second acquisition unit is configured to acquire an attribute recognition pre-trained model and item labeling information for the second target item, wherein the attribute recognition pre-trained model is generated based on the method described in any one of claims 1-3; The determining unit is configured to determine the task information corresponding to the downstream task, wherein the downstream task is a task associated with the item attribute; The model structure adjustment unit is configured to perform a first model structure adjustment on the attribute recognition pre-trained model in response to determining that the task information is attribute recognition task information, so as to generate an adjusted attribute recognition model. The second training unit is configured to train the adjusted attribute recognition model based on the item labeling information to obtain the attribute recognition model.

11. A device for generating full-scale item attribute information, comprising: The third acquisition unit is configured to acquire third item information for the third target item; The fourth input unit is configured to input the third item information into a pre-trained attribute recognition model to output full item attribute information for the third target item, wherein the attribute recognition model is generated based on the method described in any one of claims 4-7.

12. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-8.

13. A computer-readable medium having a computer program stored thereon, wherein, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.

14. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-8.

Images