CN-117036858-B - Biological object identification method, device, equipment and storage medium

CN117036858BCN 117036858 BCN117036858 BCN 117036858BCN-117036858-B

Abstract

The application discloses a method, a device, equipment and a storage medium for identifying a biological object, wherein the method comprises the steps of identifying the target biological object through a biological identification model, and determining an identification result, wherein a biological sample image is used as a training sample in the establishment process of the biological identification model, the biological identification model is obtained through training under a class balance loss function, the class balance loss function comprises a first addition and a second addition, the first addition is used for balancing training loss corresponding to the first addition in the model training process, and the second addition is used for balancing training loss corresponding to the second addition in the model training process. Therefore, a class equalization loss function is introduced in the model training process, so that the loss of a large number of classes is reduced, the loss of a small number of classes is increased, the contribution of different classes to the loss is relatively balanced, the influence of the model on the performance due to the long tail of data is improved, and the misjudgment rate and the rejection rate of the model are reduced.

Inventors

YANG QI
CHEN SHUKAI

Assignees

熵基科技股份有限公司

Dates

Publication Date: 20260505
Application Date: 20230811

Claims (6)

1. A method of identifying a biological object, comprising: Identifying a target biological object through a pre-established biological identification model, and determining an identification result of the target biological object; The establishment process of the biological recognition model comprises the following steps: Acquiring a plurality of biological images; preprocessing the multiple biological images to obtain multiple biological sample images; Training the multiple biological sample images as training samples under a class balance loss function to obtain a biological recognition model, wherein the class balance loss function comprises a first balance addition and a second balance addition, the first balance addition is used for balancing training loss of the biological recognition model corresponding to the first balance addition in the training process, and the second balance addition is used for balancing training loss of the biological recognition model corresponding to the second balance addition in the training process; the class equalization loss function is a class equalization first loss function, a class equalization second loss function or a class equalization third loss function, and the class equalization first loss function is: Wherein, the For the preset positive type label to be used, S is a preset negative label, s is a preset biological characteristic scale, For marking with the positive type label A first included angle between a feature vector of a biometric feature and a weight vector corresponding to the biometric feature, For marking a second included angle between the feature vector of the biological feature of the negative class label j and the weight vector corresponding to the biological feature, m is the punishment angle interval of the first included angle, For marking with the positive type label Is used to determine the number of images of the biological sample, For the number of biological sample images marked with the negative class label j, For the first equalization addition, Adding for the second equalization; The class equalization second loss function is: 。
2. the method of claim 1, wherein the class-balancing third loss function is: Wherein, the To replace the angle interval value of the penalty angle interval, Additional interval values for adjusting the angular interval value.
3. A biological object recognition apparatus, comprising: The model identification unit is used for identifying the target biological object through a pre-established biological identification model and determining the identification result of the target biological object; a biological image acquisition unit configured to acquire a plurality of biological images; the image preprocessing unit is used for preprocessing the biological images to obtain a plurality of biological sample images; The model training unit is used for training the plurality of biological sample images to obtain a biological identification model by taking the biological sample images as training samples under a class balance loss function, wherein the class balance loss function comprises a first balance addition and a second balance addition, the first balance addition is used for balancing the training loss of the biological identification model corresponding to the first balance addition in the training process, and the second balance addition is used for balancing the training loss of the biological identification model corresponding to the second balance addition in the training process; the class equalization loss function is a class equalization first loss function, a class equalization second loss function or a class equalization third loss function, and the class equalization first loss function is: Wherein, the For the preset positive type label to be used, S is a preset negative label, s is a preset biological characteristic scale, For marking with the positive type label A first included angle between a feature vector of a biometric feature and a weight vector corresponding to the biometric feature, For marking a second included angle between the feature vector of the biological feature of the negative class label j and the weight vector corresponding to the biological feature, m is the punishment angle interval of the first included angle, For marking with the positive type label Is used to determine the number of images of the biological sample, For the number of biological sample images marked with the negative class label j, For the first equalization addition, Adding for the second equalization; The class equalization second loss function is: 。
4. the apparatus of claim 3, wherein the class-equalization third loss function is: Wherein, the To replace the angle interval value of the penalty angle interval, Additional interval values for adjusting the angular interval value.
5. An identification device for a biological object, comprising a memory and a processor; the memory is used for storing programs; the processor for executing the program to implement the respective steps of the biological object identification method according to any one of claims 1 and 2.
6. A storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the method of identifying biological objects according to any one of claims 1 and 2.

Description

Biological object identification method, device, equipment and storage medium Technical Field The application relates to the technical field of intelligent security, in particular to a method, a device, equipment and a storage medium for identifying biological objects. Background With the continuous development of information technology, information mining is interesting for a plurality of researchers, and useful information is mined from massive information. For different biological objects, the biological objects have corresponding unique characteristics, such as facial features of people, and researchers can determine a specific object by extracting characteristic information from a face image or complicated information of the face image. The existing entrance guard machine can input face features in advance, and the target object is identified through a face recognition model so as to judge whether the target object passes or not. However, when the number of IDs and the number of the grade base are increased, the prior gate inhibition machine can have misjudgment and refusal, which indicates that the misjudgment rate and refusal rate of the biological recognition model applied to the gate inhibition machine are higher. How to train a biological recognition model with low misjudgment rate and low rejection rate so as to improve the capability of the biological recognition model for recognizing biological objects is a problem needing attention. Disclosure of Invention In view of the foregoing, the present application is directed to a method, apparatus, device, and storage medium for identifying a biological object, so as to enhance the ability of a biological identification model to identify the biological object. In order to achieve the above object, the following specific solutions are proposed: a method of identifying a biological object, comprising: Identifying a target biological object through a pre-established biological identification model, and determining an identification result of the target biological object; The establishment process of the biological recognition model comprises the following steps: Acquiring a plurality of biological images; preprocessing the multiple biological images to obtain multiple biological sample images; The method comprises the steps of taking a plurality of biological sample images as training samples, and training under a class balance loss function to obtain a biological identification model, wherein the class balance loss function comprises a first balance addition and a second balance addition, the first balance addition is used for balancing training loss of the biological identification model corresponding to the first balance addition in the training process, and the second balance addition is used for balancing training loss of the biological identification model corresponding to the second balance addition in the training process. Optionally, the class equalization loss function is a class equalization first loss function, a class equalization second loss function, or a class equalization third loss function, and the class equalization first loss function is: Wherein yi is a preset positive type label, j is a preset negative type label, s is a preset biological feature scale, θ yi is a first included angle between a feature vector marked with a biological feature of the positive type label yi and a weight vector corresponding to the biological feature, θ j is a second included angle between a feature vector marked with a biological feature of the negative type label j and a weight vector corresponding to the biological feature, m is a punishment angle interval of the first included angle, n yi is the number of biological sample images marked with the positive type label yi, n j is the number of biological sample images marked with the negative type label j, ln (n yi) is the first balanced addition, and ln (n j) is the second balanced addition. Optionally, the class equalization second loss function is: optionally, the class equalization third loss function is: where g angle is an angular interval value that replaces the penalty angular interval, and g add is an additional interval value that adjusts the angular interval value. An identification device of a biological object, comprising: The model identification unit is used for identifying the target biological object through a pre-established biological identification model and determining the identification result of the target biological object; a biological image acquisition unit configured to acquire a plurality of biological images; the image preprocessing unit is used for preprocessing the biological images to obtain a plurality of biological sample images; the model training unit is used for training the plurality of biological sample images to obtain a biological identification model by taking the biological sample images as training samples under a class balance loss function, wherein the class