CN-122024288-A - Biological feature recognition method, device, storage medium and computer equipment

CN122024288ACN 122024288 ACN122024288 ACN 122024288ACN-122024288-A

Abstract

The biological characteristic recognition method, the device, the storage medium and the computer equipment provided by the application are used for determining the characteristic recognition model trained in advance, and the characteristic space distribution of the model can be optimized by training the model by introducing the objective loss function consisting of the diversity promoting loss and the characteristic recognition loss in the training process of the characteristic recognition model. Further, the diversity promoting loss promotes the inter-class characteristics in the characteristic recognition model to be separated from each other and clear in boundary by minimizing the similarity of different types of characteristics and applying constraint punishment of the angle layers of the different types of characteristics, so that the model can learn more discriminative characteristics to improve the recognition capability. Finally, inputting the target biological characteristics into the characteristic recognition model, wherein the model can output more reliable characteristic recognition results by means of the learned strong discriminant characteristic space, so that the occurrence probability of mismatching and false rejection is fundamentally reduced.

Inventors

YANG QI
CHEN SHUKAI
ZHONG CHONGLIANG

Assignees

厦门熵基科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260224

Claims (10)

1. A method of biometric identification, the method comprising: acquiring biological characteristics to be identified; Preprocessing the biological characteristics to obtain target biological characteristics; determining a pre-trained feature recognition model, wherein the feature recognition model is obtained by training a target loss function formed by diversity promoting loss and feature recognition loss, and the diversity promoting loss promotes feature separation among feature spaces in the feature recognition model by minimizing similarity of different types of features and angle constraint penalty of the different types of features; And inputting the target biological characteristics into the characteristic recognition model to obtain a characteristic recognition result.
2. The method of claim 1, wherein the training process of the feature recognition model comprises: acquiring a biological feature training sample set; preprocessing each biological characteristic sample in the biological characteristic training sample set to obtain training sample characteristics; Constructing an initial feature recognition model, taking the minimized target loss function as a training target, and performing iterative training on the initial feature recognition model based on the training sample features and sample labels thereof, wherein the target loss function is a weighted sum of diversity promotion loss and feature recognition loss; and when the iterative training is carried out until the preset training conditions are met, determining the initial feature recognition model as a feature recognition model.
3. The biometric identification method according to claim 1 or 2, wherein the diversity promoting loss calculation process includes: Acquiring training sample characteristics; Calculating the similarity between each feature vector in the training sample features to form a similarity matrix; Determining a mask matrix, and removing self-similarity items in the similarity matrix based on the mask matrix to obtain basic loss items; And limiting angles among the feature vectors in the training sample features to form angle penalty terms, and generating diversity promotion losses according to the basic loss terms and the angle penalty terms.
4. The biometric identification method according to claim 1 or 2, wherein the calculation process of the feature identification loss includes: Acquiring a training sample feature and a weight vector of a target class corresponding to each feature vector in the training sample feature; Respectively carrying out normalization processing on each feature vector and each weight vector, and calculating an included angle cosine value between each normalized feature vector and the corresponding normalized weight vector; adding a preset angle interval to each included angle cosine value to obtain a plurality of corrected cosine values; and generating feature recognition loss according to each feature vector and the corresponding modified cosine value.
5. The biometric identification method according to claim 1 or 2, wherein the objective loss function comprises: In the formula, The target loss function is represented as a function of the target loss, 、 As the weight of the material to be weighed, Representing a loss of feature recognition, Representing a loss of promotion of the diversity, The similarity matrix is represented by a matrix of similarity, Representing a mask matrix.
6. A method of biometric identification as in claim 3 wherein said calculating the similarity between each of the feature vectors in the training sample features to form a similarity matrix comprises: determining a feature vector matrix based on each feature vector in the training sample features; And carrying out dot product on the eigenvector matrix and the transposed matrix thereof to obtain an original similarity matrix, and carrying out translation processing on each element of the original similarity matrix to obtain a similarity matrix.
7. The method of claim 1, wherein the preprocessing the biometric feature to obtain a target biometric feature comprises: Noise removing treatment is carried out on the biological characteristics, environmental interference is eliminated, equipment noise is collected, and denoising biological characteristics are obtained; and extracting the characteristics of the denoising biological characteristics to obtain target biological characteristics.
8. A biometric identification device, the device comprising: the characteristic acquisition module is used for acquiring biological characteristics to be identified; The pretreatment module is used for carrying out pretreatment on the biological characteristics to obtain target biological characteristics; The model determining module is used for determining a pre-trained feature recognition model, the feature recognition model is obtained by training a target loss function formed by diversity promoting loss and feature recognition loss, and the diversity promoting loss promotes feature separation among feature spaces in the feature recognition model by minimizing similarity of different types of features and angle constraint penalty of the different types of features; And the feature recognition module is used for inputting the target biological feature into the feature recognition model to obtain a feature recognition result.
9. A storage medium having stored therein computer readable instructions which, when executed by one or more processors, cause the one or more processors to perform the steps of the biometric method of any of claims 1 to 7.
10. A computer device includes one or more processors and a memory; stored in the memory are computer readable instructions which, when executed by the one or more processors, perform the steps of the biometric method as claimed in any one of claims 1 to 7.

Description

Biological feature recognition method, device, storage medium and computer equipment Technical Field The present application relates to the field of deep learning technologies, and in particular, to a method and apparatus for identifying biological features, a storage medium, and a computer device. Background In the scenes of attendance checking, entrance guard, identity verification and the like in the security field, the biological feature recognition technology is widely applied, and the existing attendance gate inhibition machine has a certain recognition rate and a lower misjudgment rate, but with the continuous expansion of the number of IDs and the scale of a registered base in the actual application, the problems of misjudgment, refusal recognition and the like easily occur in the recognition process, and the use experience of users and the reliability of a recognition system are affected. At present, deep learning models are widely applied in the field of biological feature recognition, and the models can realize recognition with certain precision in a conventional scene. However, when the method is used in a large-scale base scene, the generalization capability of identification is difficult to be effectively improved, the defect of insufficient accuracy of feature identification is commonly existed, and the requirements of low misjudgment rate and low rejection rate in practical application cannot be met. Disclosure of Invention The application aims to at least solve one of the technical defects, and particularly aims to solve the technical defects that the identification generalization capability is difficult to be effectively improved and the feature identification accuracy is insufficient in the prior art when the large-scale base scene is faced. In a first aspect, the present application provides a method of biometric identification, the method comprising: acquiring biological characteristics to be identified; Preprocessing the biological characteristics to obtain target biological characteristics; determining a pre-trained feature recognition model, wherein the feature recognition model is obtained by training a target loss function formed by diversity promoting loss and feature recognition loss, and the diversity promoting loss promotes feature separation among feature spaces in the feature recognition model by minimizing similarity of different types of features and angle constraint penalty of the different types of features; And inputting the target biological characteristics into the characteristic recognition model to obtain a characteristic recognition result. In one embodiment, the training process of the feature recognition model includes: acquiring a biological feature training sample set; preprocessing each biological characteristic sample in the biological characteristic training sample set to obtain training sample characteristics; Constructing an initial feature recognition model, taking the minimized target loss function as a training target, and performing iterative training on the initial feature recognition model based on the training sample features and sample labels thereof, wherein the target loss function is a weighted sum of diversity promotion loss and feature recognition loss; and when the iterative training is carried out until the preset training conditions are met, determining the initial feature recognition model as a feature recognition model. In one embodiment, the diversity facilitates a loss calculation process comprising: Acquiring training sample characteristics; Calculating the similarity between each feature vector in the training sample features to form a similarity matrix; Determining a mask matrix, and removing self-similarity items in the similarity matrix based on the mask matrix to obtain basic loss items; And limiting angles among the feature vectors in the training sample features to form angle penalty terms, and generating diversity promotion losses according to the basic loss terms and the angle penalty terms. In one embodiment, the calculation process of the feature recognition loss includes: Acquiring a training sample feature and a weight vector of a target class corresponding to each feature vector in the training sample feature; Respectively carrying out normalization processing on each feature vector and each weight vector, and calculating an included angle cosine value between each normalized feature vector and the corresponding normalized weight vector; adding a preset angle interval to each included angle cosine value to obtain a plurality of corrected cosine values; and generating feature recognition loss according to each feature vector and the corresponding modified cosine value. In one embodiment, the objective loss function includes: In the formula, The target loss function is represented as a function of the target loss,、As the weight of the material to be weighed,Representing a loss of feature recognition,Representing a loss of pr