A method, apparatus, device, and storage medium for identifying biological objects.

Abstract

This application discloses a method, apparatus, device, and storage medium for identifying biological objects. The method includes: identifying a target biological object using a biometric model and determining the identification result. The biometric model is established by using biological sample images as training samples and training the model under a class balance loss function. The class balance loss function includes a first additive term and a second additive term. The first additive term is used to balance the training loss corresponding to the first additive term during model training, and the second additive term is used to balance the training loss corresponding to the second additive term during model training. It is evident that the introduction of a class balance loss function during model training reduces the loss of classes with a larger number of classes and increases the loss of classes with a smaller number of classes, thus relatively balancing the contribution of different classes to the loss. This mitigates the impact of long-tailed data on model performance and reduces the false positive and false negative rates.

Description

Technical Field

[0001] This application relates to the field of intelligent security technology, and more specifically, to a method, device, equipment, and storage medium for identifying biological objects. Background Technology

[0002] With the continuous development of information technology, information mining has attracted the interest of many researchers, who extract useful information from massive amounts of data. Different biological objects possess unique characteristics; for example, human facial features allow researchers to extract feature information from facial images or complex facial data to identify a specific object. Existing access control systems can pre-record facial features and use facial recognition models to identify target objects and determine whether they are permitted passage.

[0003] However, current access control machines experience misjudgments and rejections as the number of IDs and the database of levels increases, indicating that the biometric models applied to access control machines have high misjudgment and rejection rates.

[0004] How to train a biometric model with low false positive and low false negative rates to improve the ability of biometric models to identify biological objects is an issue that needs attention. Summary of the Invention

[0005] In view of the above problems, this application is made to provide a method, apparatus, device and storage medium for identifying biological objects, so as to improve the ability of biometric models to identify biological objects.

[0006] To achieve the above objectives, the following specific solutions are proposed:

[0007] A method for identifying biological objects includes:

[0008] The target biological object is identified by a pre-established biometric model, and the identification result of the target biological object is determined.

[0009] The process of establishing the biometric model includes:

[0010] Acquire multiple biological images;

[0011] The multiple biological images are preprocessed to obtain multiple biological sample images;

[0012] Using the multiple biological sample images as training samples, a biometric model is trained under a class balance loss function. The class balance loss function includes a first balance term and a second balance term. The first balance term is used to balance the training loss of the biometric model corresponding to the first balance term during the training process, and the second balance term is used to balance the training loss of the biometric model corresponding to the second balance term during the training process.

[0013] Optionally, the category balance loss function is a first category balance loss function, a second category balance loss function, or a third category balance loss function, wherein the first category balance loss function is:

[0014]

[0015] Where yi is the preset positive class label, j is the preset negative class label, s is the preset biometric scale, and θ is the preset biometric scale. yi θ is the first angle between the feature vector of the biometric feature labeled with the positive class label yi and the weight vector corresponding to that biometric feature. j Let m be the second angle between the feature vector of the biometric feature labeled with the negative class label j and the weight vector corresponding to the biometric feature, and m be the penalty angle interval of the first angle, and n be the second angle between the feature vector and the weight vector of the biometric feature labeled with the negative class label j. yi n represents the number of biological sample images labeled with the positive class label yi. j The number of biological sample images labeled with the negative class label j, ln(n yi ) represents the first equilibrium addition term, ln(n) j ) is the second equilibrium addition term.

[0016] Optionally, the second loss function for class balance is:

[0017]

[0018] Optionally, the third loss function for class balance is:

[0019]

[0020] Among them, g angle To replace the angular interval value of the penalty angular interval, g add Additional interval values ​​are used to adjust the angular interval values.

[0021] A biological object identification device, comprising:

[0022] The model recognition unit is used to identify the target biological object through a pre-established biometric model and determine the recognition result of the target biological object;

[0023] Biological image acquisition unit, used to acquire multiple biological images;

[0024] An image preprocessing unit is used to preprocess the plurality of biological images to obtain a plurality of biological sample images;

[0025] The model training unit is used to train a biometric model using the plurality of biological sample images as training samples under a class balance loss function. The class balance loss function includes a first balance term and a second balance term. The first balance term is used to balance the training loss of the biometric model corresponding to the first balance term during the training process, and the second balance term is used to balance the training loss of the biometric model corresponding to the second balance term during the training process.

[0026] Optionally, the category balance loss function is a first category balance loss function, a second category balance loss function, or a third category balance loss function, wherein the first category balance loss function is:

[0027]

[0028] Where yi is the preset positive class label, j is the preset negative class label, s is the preset biometric scale, and θ is the preset biometric scale. yi θ is the first angle between the feature vector of the biometric feature labeled with the positive class label yi and the weight vector corresponding to that biometric feature. j Let m be the second angle between the feature vector of the biometric feature labeled with the negative class label j and the weight vector corresponding to the biometric feature, and m be the penalty angle interval of the first angle, and n be the second angle between the feature vector and the weight vector of the biometric feature labeled with the negative class label j. yi n represents the number of biological sample images labeled with the positive class label yi. j The number of biological sample images labeled with the negative class label j, ln(n yi ) represents the first equilibrium addition term, ln(n) j ) is the second equilibrium addition term.

[0029] Optionally, the second loss function for class balance is:

[0030]

[0031] Optionally, the third loss function for class balance is:

[0032]

[0033] Among them, g angle To replace the angular interval value of the penalty angular interval, g add Additional interval values ​​are used to adjust the angular interval values.

[0034] A device for identifying biological objects, including a memory and a processor;

[0035] The memory is used to store programs;

[0036] The processor is used to execute the program to implement the various steps of the biological object identification method as described above.

[0037] A storage medium having a computer program stored thereon, which, when executed by a processor, implements the various steps of the biological object identification method as described above.

[0038] Using the above technical solution, this application identifies target biological objects through a pre-established biometric model and determines the identification result of the target biological object. The biometric model is established by acquiring multiple biological images, preprocessing these images to obtain multiple biological sample images, and using these biological sample images as training samples. The biometric model is trained under a class balance loss function, which includes a first balance term and a second balance term. The first balance term balances the training loss of the biometric model corresponding to the first balance term during training, and the second balance term balances the training loss of the biometric model corresponding to the second balance term during training. Therefore, by introducing a class balance loss function containing both the first and second balance terms during the training of the biometric model, the loss of classes with a larger number of classes is reduced, while the loss of classes with a smaller number of classes is increased. This results in a relatively balanced contribution of different classes to the loss, mitigating the impact of long-tail data on the performance of the biometric model and reducing the false positive and false negative rates. Attached Figure Description

[0039] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0040] Figure 1 A schematic diagram illustrating the process of establishing a biometric model provided in the embodiments of this application;

[0041] Figure 2 This is a schematic diagram of a device structure for recognizing biological objects provided in an embodiment of this application;

[0042] Figure 3 This is a schematic diagram of the structure of a device for recognizing biological objects, provided as an embodiment of this application. Detailed Implementation

[0043] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0044] The proposed solution can be implemented based on a terminal with data processing capabilities, such as a computer, server, cloud, access control system, etc.

[0045] Next, the biological object identification method of this application can identify the target biological object by using a pre-established biometric model to determine the identification result of the target biological object.

[0046] Among them, the pre-established biometric model can be established as follows: Figure 1 As shown, the establishment process may include the following steps:

[0047] Step S110: Acquire multiple biological images.

[0048] Specifically, the biological image can be a facial image, and multiple biological images can be obtained from a pre-set database. It should be noted that the implementation of this invention constitutes legal use.

[0049] Step S120: Preprocess multiple biological images to obtain multiple biological sample images.

[0050] Understandably, the training data used for model training needs to conform to the data format input to the biometric model. Since the data formats of multiple biological images may be inconsistent and may not conform to the data format input to the biometric model, it is necessary to preprocess multiple biological images to obtain multiple biological sample images, all of which must conform to the data format input to the biometric model.

[0051] For example, if the data input to the biometric model needs to be in the form of four-dimensional tensor data, then multiple biological images can be preprocessed to convert each biological image into a four-dimensional tensor. The specific preprocessing process could be: cropping the biological image size, setting input channels, defining the batch of input biological images, and constructing the four-dimensional tensor.

[0052] Step S130: Using multiple biological sample images as training samples, a biometric recognition model is trained under the category balance loss function.

[0053] Specifically, the class balance loss function can include a first balance addition term and a second balance addition term.

[0054] The first balancing term can be used to balance the training loss of the biometric model during training, corresponding to the first balancing term. The second balancing term can be used to balance the training loss of the biometric model during training, corresponding to the second balancing term.

[0055] Understandably, the training process of a biometric model includes the step of classifying each biometric image. Specifically, each biometric image can be labeled with a positive or negative class. The positive class can represent the category pre-specified according to the sampling rules in a binary classification problem, and the negative class can represent the category complementary to the positive class. Since there may be a class imbalance among multiple biometric images, the training loss of the two classes can be balanced by the first and second balancing terms, respectively, so that the training loss of the class with more data is smaller and the training loss of the class with fewer data is larger.

[0056] For example, taking MobileFaceNet as an example, when training the model using a small database and without introducing a class balance loss function during training, the resulting Receiver Operator Characteristic Curve (ROC) shows that when the False Accept Rate (FAR) is 1e -3 At that time, the false rejection rate (FRR) was 0.5729%, and when FRR = 1e -4 When FRR = 1.8582%, when FAR = 1e -5 When FRR = 5.1304%, when FAR = 1e -6 When FAR = 1, FRR = 9.4263%; when FAR = 0, FRR = 14.0697%. Under the category balance loss function, the ROC curve of the trained biometric model shows that when FAR = 1 / e... -3 When FRR = 0.5232%, when FAR = 1e -4 When FRR = 1.7307%, and when FAR = 1e -5 When FRR = 4.2704%, and FAR = 1e -6 When FAR = 0, FRR = 6.8383%; when FAR = 0, FRR = 10.8082%. (This is followed by a continuation of the previous sentence, which is incomplete and requires further context.) -5 and FAR = 1e -6As can be seen, the FRR of the biometric model is reduced compared to MobileFaceNet. Therefore, the biometric model trained under the class balance loss function can improve the biometric performance.

[0057] The biological object recognition method provided in this embodiment identifies target biological objects through a pre-established biometric model and determines the recognition result of the target biological object. The biometric model establishment process involves acquiring multiple biological images, preprocessing these images to obtain multiple biological sample images, and using these biological sample images as training samples to train the biometric model under a class balance loss function. The class balance loss function includes a first balance term and a second balance term. The first balance term is used to balance the training loss of the biometric model corresponding to the first balance term during training, and the second balance term is used to balance the training loss of the biometric model corresponding to the second balance term during training. Therefore, by introducing a class balance loss function containing both the first and second balance terms during the training of the biometric model, the loss of classes with a larger number of classes is reduced, while the loss of classes with a smaller number of classes is increased. This results in a relatively balanced contribution of different classes to the loss, mitigating the impact of long-tail data on the performance of the biometric model and reducing the false positive and false negative rates.

[0058] In some embodiments of this application, the category balance loss function mentioned in the above embodiments is described. Specifically, the category balance loss function may be a first category balance loss function, a second category balance loss function, or a third category balance loss function.

[0059] The first loss function for class balance can be:

[0060]

[0061] Where yi is the preset positive class label, j is the preset negative class label, s is the preset biometric scale, and θ is the preset biometric scale. yi θ is the first angle between the feature vector of the biometric feature labeled with the positive class label yi and the weight vector corresponding to that biometric feature. j Let m be the second angle between the feature vector of the biometric feature labeled with the negative class label j and the weight vector corresponding to the biometric feature, and m be the penalty angle interval of the first angle, and n be the second angle between the feature vector and the weight vector of the biometric feature labeled with the negative class label j. yi n represents the number of biological sample images labeled with the positive class label yi. j The number of biological sample images labeled with the negative class label j, ln(n yi ) represents the first equilibrium addition term, ln(n) j) is the second equilibrium addition term.

[0062] Specifically, the first loss function for class balance can add a first balance term and a second balance term to each exponential function in the first loss function.

[0063] The first loss function can be:

[0064]

[0065] The second loss function for class balance can be:

[0066]

[0067] Specifically, the category balance second loss function can add a first balance term and a second balance term to each exponential function in the second loss function.

[0068] The second loss function can be:

[0069]

[0070] The third loss function for class balance can be:

[0071]

[0072] Among them, g angle To replace the angular interval value of the penalty angular interval, g add Additional interval values ​​are used to adjust the angular interval values. When the image quality of the biological sample image is high, g angle and g add g can be used to emphasize hard samples, but when the image quality of biological sample images is low, angle and g add This is not used to emphasize hard samples.

[0073] Specifically, the third loss function for class balance can add a first balance term and a second balance term to each exponential function in the third loss function.

[0074] The third loss function can be:

[0075]

[0076] The apparatus for recognizing biological objects provided in the embodiments of this application is described below. The apparatus for recognizing biological objects described below can be referred to in correspondence with the method for recognizing biological objects described above.

[0077] See Figure 2 , Figure 2This is a schematic diagram of a device for recognizing biological objects disclosed in an embodiment of this application.

[0078] like Figure 2 As shown, the device may include:

[0079] The model recognition unit 11 is used to identify the target biological object through a pre-established biometric model and determine the recognition result of the target biological object;

[0080] Biological image acquisition unit 12 is used to acquire multiple biological images;

[0081] Image preprocessing unit 13 is used to preprocess the plurality of biological images to obtain a plurality of biological sample images;

[0082] The model training unit 14 is used to train a biometric model using the plurality of biological sample images as training samples under a class balance loss function. The class balance loss function includes a first balance term and a second balance term. The first balance term is used to balance the training loss of the biometric model corresponding to the first balance term during the training process, and the second balance term is used to balance the training loss of the biometric model corresponding to the second balance term during the training process.

[0083] Optionally, the category balance loss function is a first category balance loss function, a second category balance loss function, or a third category balance loss function, wherein the first category balance loss function is:

[0084]

[0085] Where yi is the preset positive class label, j is the preset negative class label, s is the preset biometric scale, and θ is the preset biometric scale. yi θ is the first angle between the feature vector of the biometric feature labeled with the positive class label yi and the weight vector corresponding to that biometric feature. j Let m be the second angle between the feature vector of the biometric feature labeled with the negative class label j and the weight vector corresponding to the biometric feature, and m be the penalty angle interval of the first angle, and n be the second angle between the feature vector and the weight vector of the biometric feature labeled with the negative class label j. yi n represents the number of biological sample images labeled with the positive class label yi. j The number of biological sample images labeled with the negative class label j, ln(n yi ) represents the first equilibrium addition term, ln(n) j ) is the second equilibrium addition term.

[0086] Optionally, the second loss function for class balance is:

[0087]

[0088] Optionally, the third loss function for class balance is:

[0089]

[0090] Among them, g angle To replace the angular interval value of the penalty angular interval, g add Additional interval values ​​are used to adjust the angular interval values.

[0091] The biometric identification device provided in this application can be applied to devices for identifying biometric objects, such as terminals: mobile phones, computers, access control systems, etc. Optionally, Figure 3 A hardware block diagram of a device for identifying biological objects is shown, with reference to... Figure 3 The hardware structure of a device for identifying biological objects may include: at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4;

[0092] In this embodiment of the application, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4;

[0093] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0094] Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0095] The memory stores a program, which the processor can call. The program is used for:

[0096] The target biological object is identified by a pre-established biometric model, and the identification result of the target biological object is determined.

[0097] The process of establishing the biometric model includes:

[0098] Acquire multiple biological images;

[0099] The multiple biological images are preprocessed to obtain multiple biological sample images;

[0100] Using the multiple biological sample images as training samples, a biometric model is trained under a class balance loss function. The class balance loss function includes a first balance term and a second balance term. The first balance term is used to balance the training loss of the biometric model corresponding to the first balance term during the training process, and the second balance term is used to balance the training loss of the biometric model corresponding to the second balance term during the training process.

[0101] Optionally, the refined and extended functions of the program can be found in the description above.

[0102] This application embodiment also provides a storage medium that can store a program suitable for execution by a processor, the program being used for:

[0103] The target biological object is identified by a pre-established biometric model, and the identification result of the target biological object is determined.

[0104] The process of establishing the biometric model includes:

[0105] Acquire multiple biological images;

[0106] The multiple biological images are preprocessed to obtain multiple biological sample images;

[0107] Using the multiple biological sample images as training samples, a biometric model is trained under a class balance loss function. The class balance loss function includes a first balance term and a second balance term. The first balance term is used to balance the training loss of the biometric model corresponding to the first balance term during the training process, and the second balance term is used to balance the training loss of the biometric model corresponding to the second balance term during the training process.

[0108] Optionally, the refined and extended functions of the program can be found in the description above.

[0109] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0110] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0111] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying biological objects, characterized in that, include: The target biological object is identified by a pre-established biometric model, and the identification result of the target biological object is determined. The process of establishing the biometric model includes: Acquire multiple biological images; The multiple biological images are preprocessed to obtain multiple biological sample images; Using the multiple biological sample images as training samples, a biometric model is trained under a class balance loss function. The class balance loss function includes a first balance term and a second balance term. The first balance term is used to balance the training loss of the biometric model during the training process corresponding to the first balance term, and the second balance term is used to balance the training loss of the biometric model during the training process corresponding to the second balance term. The category balance loss function is a first category balance loss function, a second category balance loss function, or a third category balance loss function, wherein the first category balance loss function is: in, For the preset positive class tags, is the preset negative class label, and s is the preset biometric scale. For those marked with the positive class label The first angle between the feature vector of a biological feature and the weight vector corresponding to that biological feature. Let m be the second angle between the feature vector of the biometric feature labeled with the negative class j and the weight vector corresponding to the biometric feature, and m be the penalty angle interval of the first angle. For those marked with the positive class label The number of biological sample images, The number of biological sample images labeled with the negative class label j, Add a term to the first equilibrium. Add a term to the second equilibrium; The second loss function for category balance is: 。 2. The method according to claim 1, characterized in that, The third loss function for category balance is: in, To replace the angular interval value of the penalty angular interval, Additional interval values ​​are used to adjust the angular interval values.

3. A biological object identification device, characterized in that, include: The model recognition unit is used to identify the target biological object through a pre-established biometric model and determine the recognition result of the target biological object; Biological image acquisition unit, used to acquire multiple biological images; An image preprocessing unit is used to preprocess the plurality of biological images to obtain a plurality of biological sample images; The model training unit is used to train a biometric model using the plurality of biological sample images as training samples under a class balance loss function. The class balance loss function includes a first balance term and a second balance term. The first balance term is used to balance the training loss of the biometric model corresponding to the first balance term during the training process, and the second balance term is used to balance the training loss of the biometric model corresponding to the second balance term during the training process. The category balance loss function is a first category balance loss function, a second category balance loss function, or a third category balance loss function, wherein the first category balance loss function is: in, For the preset positive class tags, is the preset negative class label, and s is the preset biometric scale. For those marked with the positive class label The first angle between the feature vector of a biological feature and the weight vector corresponding to that biological feature. Let m be the second angle between the feature vector of the biometric feature labeled with the negative class j and the weight vector corresponding to the biometric feature, and m be the penalty angle interval of the first angle. For those marked with the positive class label The number of biological sample images, The number of biological sample images labeled with the negative class label j, Add a term to the first equilibrium. Add a term to the second equilibrium; The second loss function for category balance is: 。 4. The apparatus according to claim 3, characterized in that, The third loss function for category balance is: in, To replace the angular interval value of the penalty angular interval, Additional interval values ​​are used to adjust the angular interval values.

5. A biological object identification device, characterized in that, Including memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement the various steps of the biological object identification method as described in any one of claims 1 and 2.

6. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the various steps of the biological object identification method as described in any one of claims 1 and 2.

Images