Data processing method, device, storage medium and electronic device

Abstract

The embodiments of the present invention provide a data processing method, device, computer-readable storage medium and electronic device, which relate to the technical field of data processing technology. The method includes: acquiring target data; performing a feature decomposition operation on the target data through a target channel to obtain target sub-data; based on the target data, performing a branch learning operation on the target sub-data to obtain first loss information and second loss information; based on the first loss information and the second loss information, determining the target loss information of the target data. Through the present invention, the problem of low data processing efficiency is solved, thereby achieving the effect of improving data processing efficiency.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of communications, and in particular, to a data processing method, device, storage medium and electronic device. Background Art

[0002] With the continuous development of deep learning, the role of combing and monitoring data is becoming more and more obvious.

[0003] For example, in the field of video surveillance, human attribute classification, as an essential step in surveillance video analysis, is widely used in various security fields.

[0004] However, most of the current deep learning-based data analysis mainly uses fixed classification labels for learning and analysis, which has obvious disadvantages in improving the performance of the analysis model.

[0005] For example, when using fixed classification labels for supervised learning to train a human attribute classification model, the entire model training process cannot obtain spatially specific position features for specific tasks. It only relies on the model's own learning ability and it is difficult to ensure that the feature attention is focused on a specific spatial position (for example, if a hat is worn, the feature attention should be focused on the head). This leads to model overfitting and poor generalization ability.

[0006] Or when using fixed classification labels for supervised learning to train a data stream classification model, it is difficult to ensure that the feature attention is focused on a specific data channel, resulting in inaccurate analysis results of the data stream.

[0007] There is currently no better solution to the above problems. Summary of the invention

[0008] Embodiments of the present invention provide a data processing method, device, storage medium and electronic device to at least solve the problem of low data processing efficiency in related technologies.

[0009] According to an embodiment of the present invention, there is provided a data processing method, comprising:

[0010] Get target data;

[0011] Performing a feature decomposition operation on the target data through a target channel to obtain target sub-data;

[0012] Based on the target data, a branch learning operation is performed on the target sub-data to obtain first loss information and second loss information, wherein the first loss information includes loss information between the target data and the target sub-data after the branch learning operation is performed, and the second loss information includes loss information after the target sub-data performs the branch learning operation;

[0013] Based on the first loss information and the second loss information, a learning model of the target data is determined.

[0014] In an exemplary embodiment, based on the target data, performing a branch learning operation on the target sub-data to obtain first loss information and second loss information includes:

[0015] Based on the target data, performing an attention learning operation on the target sub-data of the target area to obtain the first loss information;

[0016] A classification learning operation is performed on the target sub-data to obtain first classification information; and a pooling connection calculation is performed on the first classification information to obtain second loss information.

[0017] In an exemplary embodiment, performing an attention learning operation on the target sub-data of the target area based on the target data to obtain the first loss information includes:

[0018] Performing a convolution operation on the target sub-data to obtain attention data of the target dimension;

[0019] Performing a dimension sampling operation on the target data to obtain target dimension data of the target dimension;

[0020] Perform a first loss calculation process on the target dimension data and the attention data to obtain the first loss information.

[0021] In an exemplary embodiment, determining the learning model of the target data based on the first loss information and the second loss information includes:

[0022] Obtaining information weights of the first loss information and the second loss information;

[0023] Based on the information weight, performing loss information calculation on the first loss information and the second loss information to obtain target loss information;

[0024] Based on the target loss information, the learning model is determined.

[0025] In an exemplary embodiment, after determining the learning model of the target data based on the first loss information and the second loss information, the method further includes:

[0026] Acquire first data;

[0027] The first data is processed by inference calculation based on the learning model to obtain learning data.

[0028] In an exemplary embodiment, obtaining the first data includes:

[0029] Get initial data;

[0030] A labeling operation is performed on the initial data to obtain the first data.

[0031] According to another embodiment of the present invention, there is provided a data processing device, comprising:

[0032] A data acquisition module, used to obtain target data;

[0033] A data decomposition module, used for performing a feature decomposition operation on the target data through a target channel to obtain target sub-data;

[0034] A branch learning module, configured to perform a branch learning operation on the target sub-data based on the target data to obtain first loss information and second loss information;

[0035] A model determination module is used to determine a learning model of the target data based on the first loss information and the second loss information.

[0036] In an exemplary embodiment, the branch learning module includes:

[0037] A first learning unit is used to perform an attention learning operation on the target sub-data of the target area based on the target data to obtain the first loss information;

[0038] The second learning unit is used to perform a classification learning operation on the target sub-data to obtain first classification information; and perform a pooling connection calculation on the first classification information to obtain second loss information.

[0039] In an exemplary embodiment, the first learning unit includes:

[0040] A first convolution subunit, configured to perform a convolution operation on the target sub-data to obtain attention data of a target dimension;

[0041] A sampling subunit, configured to perform a dimension sampling operation on the target data to obtain target dimension data of the target dimension;

[0042] The first loss calculation unit is used to perform a first loss calculation process on the target dimension data and the attention data to obtain the first loss information.

[0043] According to yet another embodiment of the present invention, a computer-readable storage medium is provided, in which a computer program is stored, wherein the computer program is configured to execute the steps of any one of the above method embodiments when run.

[0044] According to yet another embodiment of the present invention, there is provided an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0045] Through the present invention, the characteristics of the target data are decomposed and branch learning is performed, thereby reducing the amount of calculation in the data analysis process. At the same time, by obtaining the target loss information, it is convenient to build a learning model, so that the learning model can analyze and process specific data, thereby improving the data processing efficiency. Therefore, the problem of low data processing efficiency of the data model can be solved, and the effect of improving data processing efficiency can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a hardware structure block diagram of a mobile terminal of a data processing method according to an embodiment of the present invention;

[0047] Figure 2 is a flow chart of a data processing method according to an embodiment of the present invention;

[0048] Figure 3 is a structural block diagram of a data processing device according to an embodiment of the present invention;

[0049] Figure 4 This is a process according to a specific embodiment of the present invention Figure 1 ;

[0050] Figure 5 This is a process according to a specific embodiment of the present invention Figure 2 . DETAILED DESCRIPTION

[0051] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings and in combination with the embodiments.

[0052] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0053] The method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 FIG. 1 is a hardware structure block diagram of a mobile terminal of a data processing method according to an embodiment of the present invention. Figure 1 As shown, the mobile terminal may include one or more ( Figure 1Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data, wherein the mobile terminal may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is for illustration only and does not limit the structure of the mobile terminal. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0054] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as a computer program corresponding to a data processing method in an embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, to implement the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the mobile terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0055] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the mobile terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0056] In this embodiment, a data processing method is provided. Figure 2 is a flow chart according to an embodiment of the present invention, such as Figure 2 As shown, the process includes the following steps:

[0057] Step S202, obtaining target data;

[0058] In this embodiment, the target data is data used for model training. The target data can be (but not limited to) traffic data in a communication network, or multimedia data in a video recognition network. Correspondingly, the target data can be (but not limited to) obtained by monitoring the network through a traffic monitoring module in a communication device such as a base station, or can be collected through multimedia sensing equipment, such as a radar camera, a visible light camera, an infrared camera, etc. The target data can be (but not limited to) obtained through Ethernet, such as wireless networks such as 3G / 4G / 5G / quantum communication networks, or wired networks such as wired transmission, or a combination of wired networks or wireless networks.

[0059] It should be noted that, to facilitate training, the target data may be (but not limited to) pre-processed data, such as data attribute labeling and classification, or may be unlabeled raw data.

[0060] Step S204, performing feature decomposition operation on the target data through the target channel to obtain target sub-data;

[0061] In this embodiment, the feature decomposition operation is performed on the target data in order to decompose the data of different regions or different attributes according to different needs, so that the computing resources can be concentrated to calculate the data of different regions or different attributes, so as to improve the data processing efficiency and the data processing accuracy.

[0062] The feature decomposition operation may be (but not limited to) segmenting, classifying or specially marking the target data according to pre-set feature data, or may be other forms of decomposition.

[0063] For example, when applied to the field of communication network monitoring, data of different time series are marked according to time series features so that data of different time series contain corresponding identification information; or data of different time series are cut and identified separately, and then re-fused; or when applied to the field of image recognition to recognize human images, the human body image is first generated into a C* H *W (multi-channel) feature map through the backbone, and then the feature map is equally divided in the channel dimension according to the number of classification tasks N to obtain N groups of sub-feature maps (such as a sub-feature map containing only the head, a sub-feature map containing only the arms, etc.), where each group of features corresponds to a task, that is, the feature map is divided into several sub-feature maps according to the number of areas to be identified.

[0064] Step S206, based on the target data, performing a branch learning operation on the target sub-data to obtain first loss information and second loss information, wherein the first loss information includes loss information between the target data and the target sub-data after the branch learning operation is performed, and the second loss information includes loss information after the branch learning operation is performed on the target sub-data;

[0065] In this embodiment, branch learning operations are performed on the target sub-data respectively in order to obtain loss information in different ways, so that a suitable learning model can be constructed according to the loss information later, so that the learning model can accurately perform deep learning processing on the data.

[0066] Among them, the branch learning operation includes (but is not limited to) performing convolution calculation on the target sub-data and then performing loss calculation, or it can be joint target data and then performing loss calculation, or it can be convolution after joint target data and then performing loss calculation, or it can be other forms of learning operations; the first loss information and the second loss information can (but is not limited to) include information such as the loss value of the target sub-data after the branch learning operation, the loss value storage area, etc.

[0067] For example, when applied to the field of image processing, the sub-task feature map can be divided into two branches: the attention branch and the classification branch. The attention branch is responsible for learning the spatial position of the task part, which is used to perform a 1*1 convolution calculation on the sub-feature map to obtain a 1*H*W attention map, and then downsample the mask of the corresponding part to 1*H*W and calculate the loss with the attention map. The other classification is responsible for learning the task classification, and after global average pooling and full connection, the loss is calculated with the task classification label.

[0068] Step S208: determining target loss information of the target data based on the first loss information and the second loss information.

[0069] In this embodiment, the target loss information is determined in order to balance the loss function and avoid the model learning being biased towards one branch in the branch learning operation, thereby improving the accuracy of the loss information, and then being able to build a more accurate learning model and use the learning model for accurate data processing.

[0070] The target loss information may include (but is not limited to) loss function, loss value and other information.

[0071] For example, the attention loss and classification loss of each task are added according to a certain weight ratio to obtain the final loss.

[0072] Through the above steps, the characteristics of the target data are decomposed and branch learning is performed, thereby reducing the amount of computation in the data analysis process. At the same time, by obtaining the target loss information, it is convenient to build a learning model, so that the learning model can analyze and process specific data, thereby improving data processing efficiency, solving the problem of low data processing efficiency, and improving data processing efficiency.

[0073] The execution subject of the above steps may be a base station, a terminal, etc., but is not limited thereto.

[0074] In an optional embodiment, based on the target data, performing a branch learning operation on the target sub-data to obtain the first loss information and the second loss information includes:

[0075] Step S2062, based on the target data, performing an attention learning operation on the target sub-data of the target area to obtain first loss information;

[0076] Step S2064, performing a classification learning operation on the target sub-data to obtain first classification information; performing a pooling connection calculation on the first classification information to obtain second loss information.

[0077] In this embodiment, the attention learning operation is performed on the target data to enable the learning model to accurately learn data in a specific area or a specific type, thereby improving the recognition ability of data in a specific area or a specific type, and the classification learning operation is performed to balance the loss deviation caused by learning data in a specific area or a specific type, so that the data-specific classification can be accurately predicted in the data inference stage, thereby improving the recognition accuracy of the learning model.

[0078] The target area can be (but not limited to) a data storage area, or a certain area in an image (such as the head, arm, backpack, etc. of a human image); the attention learning operation includes (but not limited to) sequentially performing convolution calculations and loss calculations on a specific area or a specific type of data, and the pooling connection calculation can be (but not limited to) performing global average pooling and full connection operations on the target sub-data. It should be noted that the full connection operation is an operation that converts a feature vector into a final classification result, and its mathematical expression is a matrix multiplication.

[0079] In an optional embodiment, based on the target data, performing an attention learning operation on the target sub-data of the target area to obtain the first loss information includes:

[0080] Step S20622, performing a convolution operation on the target sub-data to obtain attention data of the target dimension;

[0081] Step S20624, performing a dimension sampling operation on the target data to obtain target dimension data of the target dimension;

[0082] Step S20626, perform a first loss calculation process on the target dimension data and attention data to obtain first loss information.

[0083] In this embodiment, the convolution operation is performed on the target sub-data to adjust the dimension of the target sub-data. Correspondingly, the sampling operation is performed on the target data to make the dimension of the target data the same as that of the attention data, thereby facilitating calculation.

[0084] Among them, when it is used in the field of image processing, the target data can be the data after mask processing, that is, the dimension sampling operation is performed on the masked data, so as to improve the recognition ability of the learning model.

[0085] For example, when applied to the image field, the target sub-feature map is first subjected to a 1*1 convolution to obtain a 1*H*W attention map, and the mask of the corresponding part is downsampled to 1*H*W before the loss is calculated with the attention map.

[0086] In an optional embodiment, determining the target loss information of the target data based on the first loss information and the second loss information includes:

[0087] Step S2082, obtaining information weights of the first loss information and the second loss information;

[0088] Step S2084: Based on the information weight, loss information is calculated for the first loss information and the second loss information to obtain target loss information.

[0089] In this embodiment, the loss information of the first loss information and the second loss information is calculated in order to balance the first loss information and the second loss information to avoid the calculation result having an excessively large weight and affecting the recognition ability of the learning model.

[0090] The information weight may be (but is not limited to) the weight of the first loss information and the second loss information in the calculation process; the target loss information includes (but is not limited to) target loss value, dimension of loss information and other information.

[0091] For example, the attention loss and classification loss of each task are added according to a certain weight ratio to obtain the final loss.

[0092] In an optional embodiment, after determining the target loss information of the target data based on the first loss information and the second loss information, the method further includes:

[0093] Step S2010, determining a learning model based on target loss information;

[0094] Step S2012: Perform inference calculation on the pre-acquired first data through the learning model to obtain learning data.

[0095] In this embodiment, the learning model is determined by the target loss information to improve the processing accuracy and efficiency of the data.

[0096] For example, the classified labeled data and the masked data can be used to perform supervised training on the learning model.

[0097] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0098] In the present embodiment, a data processing device is also provided, which is used to implement the above-mentioned embodiments and preferred implementation modes, and the descriptions that have been made will not be repeated. As used below, the term "module" can implement a combination of software and / or hardware of a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceived.

[0099] Figure 3 is a structural block diagram of a data processing device according to an embodiment of the present invention. Figure 2 As shown, the device comprises:

[0100] A data acquisition module 32, used to acquire target data;

[0101] A data decomposition module 34 is used to perform a feature decomposition operation on the target data through a target channel to obtain target sub-data;

[0102] A branch learning module 36, configured to perform a branch learning operation on the target sub-data based on the target data to obtain first loss information and second loss information;

[0103] The information determination module 38 is used to determine the target loss information of the target data based on the first loss information and the second loss information.

[0104] In an optional embodiment, the branch learning module 36 includes:

[0105] A first learning unit 362 is used to perform an attention learning operation on the target sub-data of the target area based on the target data to obtain the first loss information;

[0106] The second learning unit 364 is used to perform a classification learning operation on the target sub-data to obtain first classification information; and perform a pooling connection calculation on the first classification information to obtain second loss information.

[0107] In an optional embodiment, the first learning unit 362 includes:

[0108] A first convolution subunit 3622, configured to perform a convolution operation on the target sub-data to obtain attention data of a target dimension;

[0109] The sampling subunit 3624 is used to perform a dimension sampling operation on the target data to obtain target dimension data of the target dimension;

[0110] The first loss calculation unit 3626 is used to perform a first loss calculation process on the target dimension data and the attention data to obtain the first loss information.

[0111] In an optional embodiment, the information determination module 38 includes:

[0112] A weight acquisition unit 382, ​​configured to acquire information weights of the first loss information and the second loss information;

[0113] The target loss calculation unit 384 is used to perform loss information calculation on the first loss information and the second loss information based on the information weight to obtain target loss information.

[0114] In an optional embodiment, the device further comprises:

[0115] A model determination module 310, configured to determine a learning model based on the target loss information after determining the target loss information of the target data based on the first loss information and the second loss information;

[0116] The inference calculation module 312 is used to perform inference calculation processing on the pre-acquired first data through the learning model to obtain learning data.

[0117] In an optional embodiment, the device further comprises:

[0118] The initial data collection module 314 is used to collect initial data before performing inference calculation processing on the pre-acquired first data through the learning model to obtain the learning data;

[0119] The data annotation module 316 is used to perform annotation processing on the initial data to obtain first data.

[0120] It should be noted that the above modules can be implemented by software or hardware. For the latter, it can be implemented in the following ways, but not limited to: the above modules are all located in the same processor; or the above modules are located in different processors in any combination.

[0121] The present invention is described below by means of specific embodiments.

[0122] like Figure 4 As shown in the figure, when applied to the field of image recognition, model training can be performed according to the following steps:

[0123] Step S41, labeling human body attribute classification labels (for example, whether wearing a hat, top color, whether carrying a backpack, etc.);

[0124] Step S42, marking the mask corresponding to the human body attributes;

[0125] Step S43, using the mask data in step S42 to train a human semantic segmentation network model, here using the DeepLabv3+ framework.

[0126] Step S44, using the model trained in step S43 to infer the data marked in step S41, and saving the inferred mask result.

[0127] Step S45, using the classified annotation data in step S41 and the mask result in step S44 to perform supervised training on the classification model.

[0128] Among them, Figure 5 As shown in the figure, the specific process of training the human semantic segmentation network model is roughly as follows:

[0129] Step S51: The human body image generates a C* H *W feature map through backbone.

[0130] Step S52, the feature map is equally divided in the channel dimension according to the number of classification tasks N, to obtain N groups of sub-feature maps, where each group of features corresponds to a task.

[0131] Step S53, the subtask feature map is divided into two branches, the attention branch and the classification branch. The attention branch is responsible for learning the spatial position of the task part. The sub-feature map is subjected to a layer of 1*1 convolution to obtain a 1*H*W attention map. The mask of the corresponding part is downsampled to 1*H*W and then the loss is calculated with the attention map. The other classification is responsible for learning the task classification, and after global average pooling and full connection, the loss is calculated with the task classification label.

[0132] Step S54, adding the attention loss and classification loss of each task according to a certain weight ratio to obtain the final loss data. This step is to balance the loss data to avoid the model learning process being biased towards the attention branch or the classification branch, resulting in low accuracy of the learning results.

[0133] An embodiment of the present invention further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when running.

[0134] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0135] An embodiment of the present invention further provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0136] In an exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0137] For specific examples in this embodiment, reference may be made to the examples described in the above embodiments and exemplary implementation modes, and this embodiment will not be described in detail herein.

[0138] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0139] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A data processing method for image recognition, characterized in that: include: Acquire target data through communication equipment and multimedia sensor equipment; The feature map is equally divided in the channel dimension of the target channel according to the number of classification tasks, so as to perform feature decomposition operation on the target data of different regions or different attributes according to different requirements to obtain target sub-data, wherein each group of sub-feature maps in the target sub-data corresponds to one task in the classification task; Based on the target data, a branch learning operation is performed on the target sub-data to obtain first loss information and second loss information, a convolution operation is performed on the target sub-data and the target data, and an attention learning operation is performed on the target sub-data in the target area to obtain the first loss information; a classification learning operation is performed on the target sub-data to obtain first classification information; a pooling connection calculation is performed on the first classification information to obtain the second loss information, wherein the first loss information includes loss information between the target data and the target sub-data after the branch learning operation is performed, and the second loss information includes loss information after the target sub-data performs the branch learning operation; Obtain information weights of the first loss information and the second loss information, perform loss information calculation on the first loss information and the second loss information based on the information weights to obtain target loss information, determine a learning model based on the target loss information, and perform inference calculation processing on the pre-acquired first data through the learning model to obtain learning data.

2. The method according to claim 1, characterized in that The step of performing a convolution operation on the target sub-data and the target data, and performing an attention learning operation on the target sub-data of the target area to obtain the first loss information includes: Performing the convolution operation on the target sub-data to obtain attention data of the target dimension; Performing a dimension sampling operation on the target data to obtain target dimension data of the target dimension; Perform a first loss calculation process on the target dimension data and the attention data to obtain the first loss information.

3. A data processing device for image recognition, characterized in that: include: Data acquisition module, used to obtain target data through communication equipment and multimedia sensor equipment; A data decomposition module, used to equally divide the feature map on the channel dimension of the target channel according to the number of classification tasks, so as to perform feature decomposition operation on the target data of different regions or different attributes according to different requirements to obtain target sub-data, wherein each group of sub-feature maps in the target sub-data corresponds to one task in the classification task; A branch learning module is used to perform a branch learning operation on the target sub-data based on the target data to obtain first loss information and second loss information, perform a convolution operation on the target sub-data and the target data, perform an attention learning operation on the target sub-data of the target area to obtain the first loss information; perform a classification learning operation on the target sub-data to obtain first classification information; perform a pooling connection calculation on the first classification information to obtain the second loss information, wherein the first loss information includes loss information between the target data and the target sub-data after the branch learning operation is performed, and the second loss information includes loss information on the target sub-data after the branch learning operation is performed; An information determination module is used to obtain information weights of the first loss information and the second loss information, perform loss information calculation on the first loss information and the second loss information based on the information weights to obtain target loss information, determine a learning model based on the target loss information, and perform inference calculation processing on the pre-acquired first data through the learning model to obtain learning data.

4. The device according to claim 3, characterized in that The device is used to perform a convolution operation on the target sub-data and the target data in the following manner, and perform an attention learning operation on the target sub-data of the target area to obtain the first loss information: A first convolution subunit, configured to perform the convolution operation on the target sub-data to obtain attention data of a target dimension; A sampling subunit, configured to perform a dimension sampling operation on the target data to obtain target dimension data of the target dimension; The first loss calculation unit is used to perform a first loss calculation process on the target dimension data and the attention data to obtain the first loss information.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 2 when executed.

6. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 2.

Images