CN-115204287-B - Object identification method, device, equipment and storage medium

CN115204287BCN 115204287 BCN115204287 BCN 115204287BCN-115204287-B

Abstract

The invention provides an object identification method, device, equipment and storage medium, wherein the method comprises the steps of obtaining target data; the method comprises the steps of identifying an object to which target data belong based on a target identification model, wherein the target identification model is obtained by training a first training sample with a new object marked with an object mark, enabling the object mark predicted by the first training sample to be consistent with the object mark marked by the first training sample, enabling probability distribution predicted by the first training sample to approach probability distribution predicted by a basic identification model by aiming at the first training sample, enabling the basic identification model to be obtained by training a second training sample with a history object marked with the object mark, and enabling an initial target identification model to be a model capable of identifying data of the history object. According to the invention, only training data of the newly added object is adopted, so that the object to which the data of the newly added object belongs can be accurately identified, and the target identification model of the object to which the data of the historical object belongs can be accurately identified.

Inventors

JIANG JUN
WANG JIANSHE
FANG SIAN
LIU LIN
ZHAN JIANBO
XU CHENG
LIU HAIBO

Assignees

合肥讯飞数码科技有限公司

Dates

Publication Date: 20260508
Application Date: 20220713

Claims (9)

1. An object recognition method, comprising: acquiring target data, wherein the target data is to-be-identified data of an object in an object set, and the object set comprises a history object and a newly added object; identifying an object to which the target data belongs based on a target identification model obtained through pre-training, wherein: the target data is electromagnetic data of one electromagnetic individual in an electromagnetic individual set, the target recognition model is an electromagnetic individual recognition model, an object to which the target data belongs is the electromagnetic individual to which the target data belongs, or the target data is voice data of one speaker in a speaker set, the target recognition model is a speaker recognition model, and the object to which the target data belongs is the speaker to which the target data belongs; The training target of the target recognition model comprises enabling the target mark predicted by aiming at the first training sample to be consistent with the target mark marked by the first training sample, and enabling the probability distribution predicted by aiming at the first training sample to approximate to the probability distribution predicted by the basic recognition model aiming at the first training sample; The probability distribution is the probability distribution on object identifications corresponding to all objects in the object set, the basic identification model is obtained by training a second training sample marked with the object identifications of the historical objects, and the initial target identification model is a model capable of identifying the data of the historical objects.
2. The method for identifying an object according to claim 1, wherein the identifying an object to which the target data belongs based on a target identification model obtained by training in advance includes: Predicting the probability that the object identifier of the target data is each object identifier in an object identifier set based on the target identification model, wherein the object identifier set comprises object identifiers respectively corresponding to each object in the object set; and determining the object to which the target data belongs according to the probability that the object identifier of the target data is each object identifier in the object identifier set.
3. The object recognition method according to claim 1, wherein an initial object recognition model is obtained by copying the basic recognition model; The training process of the target recognition model comprises the following steps: Predicting the probability that the object identifier of the first training sample is each object identifier in an object identifier set based on a target identification model and the basic identification model respectively to obtain first probability distribution and second probability distribution, wherein the first probability distribution is the probability distribution predicted by the target identification model for the first training sample, the second probability distribution is the probability distribution predicted by the basic identification model for the first training sample, and the object identifier set comprises the object identifiers respectively corresponding to the objects in the object set; Determining a first prediction loss and a second prediction loss according to the first probability distribution, the second probability distribution and the object identification marked by the first training sample, wherein the first prediction loss can represent the similarity of the first probability distribution and the real probability distribution corresponding to the first training sample, and the second prediction loss can represent the difference of the first probability distribution and the second probability distribution, and the real probability distribution is determined by the object identification marked by the corresponding first training sample; And updating parameters of the target recognition model according to the first prediction loss and the second prediction loss.
4. The method of claim 3, wherein determining the first predicted loss and the second predicted loss based on the first probability distribution, the second probability distribution, and the object identification of the first training sample label comprises: Calculating cross entropy loss between the first probability distribution and the object mark marked by the first training sample as a first prediction loss; the distance between the first probability distribution and the second probability distribution is calculated as a second predictive loss.
5. The method of claim 4, wherein the object recognition model comprises a hidden layer and a fully connected layer; And updating parameters of a target recognition model according to the first prediction loss and the second prediction loss, wherein the method comprises the following steps: and updating parameters of the full-connection layer of the target recognition model according to the first prediction loss and the second prediction loss.
6. The method according to claim 5, wherein the updating parameters of the fully connected layer of the object recognition model according to the first prediction loss and the second prediction loss comprises: fusing the first predicted loss and the second predicted loss to obtain total predicted loss; And according to the total prediction loss, updating parameters of a full-connection layer of the target recognition model.
7. An object recognition device is characterized by comprising a data acquisition module and an object recognition module; the data acquisition module is used for acquiring target data, wherein the target data is to-be-identified data of an object in an object set, and the object set comprises a history object and a newly-added object; the object recognition module is configured to recognize an object to which the target data belongs based on a target recognition model obtained by training in advance, where: the target data is electromagnetic data of one electromagnetic individual in an electromagnetic individual set, the target recognition model is an electromagnetic individual recognition model, an object to which the target data belongs is the electromagnetic individual to which the target data belongs, or the target data is voice data of one speaker in a speaker set, the target recognition model is a speaker recognition model, and the object to which the target data belongs is the speaker to which the target data belongs; The training target of the target recognition model comprises enabling the target mark predicted by aiming at the first training sample to be consistent with the target mark marked by the first training sample, and enabling the probability distribution predicted by aiming at the first training sample to approximate to the probability distribution predicted by the basic recognition model aiming at the first training sample; The probability distribution is the probability distribution on object identifications corresponding to all objects in the object set, the basic identification model is obtained by training a second training sample marked with the object identifications of the historical objects, and the initial target identification model is a model capable of identifying the data of the historical objects.
8. An object recognition device is characterized by comprising a memory and a processor; the memory is used for storing programs; the processor is configured to execute the program to implement the steps of the object recognition method according to any one of claims 1 to 6.
9. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the object recognition method according to any one of claims 1-6.

Description

Object identification method, device, equipment and storage medium Technical Field The present invention relates to the field of object recognition technologies, and in particular, to an object recognition method, device, apparatus, and storage medium. Background In the field of object recognition, it is necessary to recognize an object to which target electromagnetic data belongs, for example, in the field of electromagnetic individual recognition, it is necessary to recognize an electromagnetic individual to which target electromagnetic data belongs, that is, data which determines which electromagnetic individual among a plurality of electromagnetic individuals the target electromagnetic data belongs. The current object recognition scheme is mainly based on a recognition scheme of a recognition model, namely, a training sample marked with an object identifier of each object is used for training in advance to obtain the recognition model, and then the object to which target data belong is recognized by using the recognition model obtained through training. In some object recognition fields, new objects may appear continuously, corresponding newly added object data may be generated, for example, in the electromagnetic individual recognition field, new electromagnetic individual deployment may appear continuously, corresponding electromagnetic data of newly added electromagnetic individuals may be generated, when new objects appear, in order to accurately recognize data of historical objects (objects before newly added objects) and accurately recognize data of newly added objects, the scheme adopted at present is that training data of existing objects and training data of newly added objects are mixed together to retrain a recognition model. However, as new subjects continue to develop, the data volume of training data continues to increase, and training models using huge volumes of training data is extremely time consuming. It can be seen that although the above solution can solve the data recognition problem of the newly added object, the efficiency of the above solution is very low. Disclosure of Invention In view of the above, the present invention provides a method, apparatus, device and storage medium for object recognition, which are used to solve the problem that when there is a new object, the training data of the existing object and the training data of the new object are mixed together to retrain the recognition model, and the training is extremely time-consuming due to the huge data volume, and the technical scheme is as follows: An object recognition method, comprising: acquiring target data, wherein the target data is to-be-identified data of an object in an object set, and the object set comprises a history object and a newly added object; identifying an object to which the target data belongs based on a target identification model obtained through pre-training, wherein: The training target of the target recognition model comprises enabling the target mark predicted by aiming at the first training sample to be consistent with the target mark marked by the first training sample, and enabling the probability distribution predicted by aiming at the first training sample to approximate to the probability distribution predicted by the basic recognition model aiming at the first training sample; The probability distribution is the probability distribution on object identifications corresponding to all objects in the object set, the basic identification model is obtained by training a second training sample marked with the object identifications of the historical objects, and the initial target identification model is a model capable of identifying the data of the historical objects. Optionally, the identifying, based on the target identification model obtained by training in advance, the object to which the target data belongs includes: Predicting the probability that the object identifier of the target data is each object identifier in an object identifier set based on the target identification model, wherein the object identifier set comprises object identifiers respectively corresponding to each object in the object set; and determining the object to which the target data belongs according to the probability that the object identifier of the target data is each object identifier in the object identifier set. Optionally, the initial target recognition model is obtained by copying the basic recognition model, and the training process of the target recognition model includes: Predicting the probability that the object identifier of the first training sample is each object identifier in an object identifier set based on a target identification model and the basic identification model respectively to obtain first probability distribution and second probability distribution, wherein the first probability distribution is the probability distribution predicted by the target identification model for the f