Human face recognition model training method and device based on MobileFaceNet

Abstract

The invention discloses a human face recognition model training method and device based on a MobileFaceNet, and the method comprises the steps: S1, obtaining a human face recognition model based on the MobileFaceNet, carrying out the improvement of a main feature extraction network in the human face recognition model, and obtaining an improved human face recognition model; s2, parameters in the improved face recognition model are optimized through an Arcfu loss function based on a dynamic boundary; and S3, training the improved face recognition model through an adaptive exponential decay learning rate strategy to obtain a final face recognition model.

Description

Technical Field

[0001] The present invention relates to the technical field of model training, and in particular to a face recognition model training method and device based on MobileFaceNet. Background Art

[0002] With the development of the Internet and the improvement of computing power in recent years, face recognition technology has been widely used in daily life and security fields, providing people with an efficient, accurate and convenient way of identity recognition and verification. Face recognition technology refers to a biometric technology based on facial images or videos, which performs identity verification or recognition by analyzing the characteristics and patterns of the face. The widespread application of face recognition technology can have a profound impact on society. It can help improve public safety, prevent crime, and can also be used for social management, demographics, etc. Therefore, it is necessary to apply face recognition technology to natural scenes.

[0003] At present, there are some excellent lightweight networks for face recognition, such as MobileFaceNet. However, for embedded devices with small capacity and very limited computing resources on the market, a lighter network is often needed. Considering the practicality of the device, the network is required to have excellent classification capabilities. Small models often mean poor feature extraction capabilities. While maintaining network recognition accuracy, reducing the size of the model and reducing the amount of calculation is a difficult point in research and development.

[0004] In view of the above technical problems, the present invention proposes a face recognition model training method and device based on MobileFaceNet. Summary of the invention

[0005] The purpose of the present invention is to provide a face recognition model training method and device based on MobileFaceNet in view of the defects of the prior art.

[0006] In order to achieve the above objectives, the present invention adopts the following technical solutions:

[0007] A face recognition model training method based on MobileFaceNet, comprising:

[0008] S1. Obtain a face recognition model based on MobileFaceNet, and improve the backbone feature extraction network in the face recognition model to obtain an improved face recognition model;

[0009] S2. Optimize the parameters of the improved face recognition model through the Arcface loss function based on dynamic boundaries;

[0010] S3. The improved face recognition model is trained through an adaptive exponential decay learning rate strategy to obtain the final face recognition model.

[0011] Furthermore, the improved backbone feature extraction network in step S1 includes: a first point-by-point convolution block and two group convolution blocks after the bottleneck layer, and a channel shuffling module is added after the two group convolution blocks.

[0012] Furthermore, in the step S2, the Arcface loss function based on the dynamic boundary is as follows: the fixed angle boundary of the Arcface loss function is changed to a dynamic angle boundary, and the parameters of the angle boundary are dynamically adjusted according to the training effect of the face recognition model.

[0013] Furthermore, the Arcface loss function based on the dynamic boundary in step S2 is expressed as:

[0014]

[0015] Where L represents the Arcface loss function based on dynamic boundaries; N represents the number of samples; e represents a constant; S represents a scaling factor; represents the angle between the feature vector of the i-th sample and the weight vector of the corresponding true category; M represents the initial angle boundary; Δm represents the parameter of the dynamic angle boundary; n represents the total number of categories; y i represents the category of the i-th sample; θ j Represents the angle between the feature vector and the j-th category weight vector.

[0016] Furthermore, the parameter Δm of the dynamic angle boundary is expressed as:

[0017]

[0018] Among them, L batch represents the loss of the current training batch; θ low represents the lower limit of the loss threshold; θ hight Indicates the upper limit of the loss threshold.

[0019] Furthermore, the learning rate strategy of adaptive exponential decay in step S3 is specifically as follows: updating the learning rate according to the training cycle length and the current training round of the face recognition model, and using the exponential decay factor to adaptively control the learning rate decrease speed to keep the face recognition model training stable.

[0020] Accordingly, a face recognition model training device based on MobileFaceNet is also provided, which is used to execute the face recognition model training method based on MobileFaceNet. The training device includes a processor, and the processor is provided with:

[0021] An improvement module is used to obtain a face recognition model based on MobileFaceNet and improve the backbone feature extraction network in the face recognition model to obtain an improved face recognition model;

[0022] An optimization module, used to optimize the parameters in the improved face recognition model through the Arcface loss function based on dynamic boundaries;

[0023] The training module is used to train the improved face recognition model through an adaptive exponential decay learning rate strategy to obtain a final face recognition model.

[0024] Furthermore, the processor is connected to a memory for storing data or instructions.

[0025] Furthermore, the processor is connected to a transmission device for transmitting data with an external device.

[0026] Furthermore, the processor is connected to an input / output device for inputting or outputting information.

[0027] Compared with the prior art, the present invention improves the network structure based on the lightweight network MobileFaceNet and designs a more lightweight network. At the same time, the Arcface loss function based on dynamic boundaries is designed for the model training effect, and the learning rate strategy of adaptive exponential decay is used to train a face recognition model with better performance, which can solve the problem that the network model is too large and cannot be deployed on small-scale embedded devices with limited computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of a face recognition model training method based on MobileFaceNet provided in Example 1;

[0029] Figure 2 This is the existing standard MobileFaceNet network structure diagram provided in Example 1. DETAILED DESCRIPTION

[0030] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0031] The purpose of the present invention is to provide a face recognition model training method and device based on MobileFaceNet in view of the defects of the prior art.

[0032] Embodiment 1

[0033] This embodiment provides a face recognition model training method based on MobileFaceNet. Figure 1 As shown, including:

[0034] S1. Obtain a face recognition model based on MobileFaceNet, and improve the backbone feature extraction network in the face recognition model to obtain an improved face recognition model;

[0035] S2. Optimize the parameters of the improved face recognition model through the Arcface loss function based on dynamic boundaries;

[0036] S3. The improved face recognition model is trained through an adaptive exponential decay learning rate strategy to obtain the final face recognition model.

[0037] Before step S1, the method also includes: obtaining a face image dataset, and dividing the face image dataset into a training set, a test set, and a validation set according to a certain ratio.

[0038] The face image dataset is obtained through database, online crawler, collection and other methods. The acquired face images are classified according to one person one ID, and each face image is processed by data cleaning, face alignment, etc., so as to obtain the face image with key points aligned and cropped; the processed face image dataset is divided into training set, test set and validation set according to the number of IDs and the ratio.

[0039] In step S1, a face recognition model based on MobileFaceNet is obtained, and the backbone feature extraction network in the face recognition model is improved to obtain an improved face recognition model.

[0040] Construct a backbone feature extraction network based on the MobileFaceNet model, and iteratively train the backbone feature extraction network using the processed training dataset.

[0041] The existing standard MobileFaceNet network consists of ordinary convolution blocks, deep convolution blocks, 15 consecutive bottleneck layers, point-by-point convolution blocks, deep convolution, and point-by-point convolution. Some adjacent bottleneck layer modules have shortcut connections; the existing standard MobileFaceNet network has a large number of parameters and computational complexity, and the network size is 4M. The existing standard MobileFaceNet network structure is shown in the figure below: Figure 2 shown.

[0042] The network designed in this embodiment is targeted at smaller embedded platforms, and its backbone network is improved on the basis of MobileFaceNet, specifically:

[0043] The improved backbone feature extraction network includes a first standard convolution block CW, a first grouped convolution block GW, 15 Bottleneck layers, a first point-by-point convolution block PW, 3 second grouped convolution blocks GW, a second point-by-point convolution block PW, and a channel shuffle module SH.

[0044] The face image size input to the improved backbone feature extraction network is 112*112, the first standard convolution block CW, the convolution kernel size of the first standard convolution block CW is 3*3, the step size s=2, and the number of output channels is 8; after the 112*112 face image passes through the first standard convolution block CW, the output size is 56*56.

[0045] The first standard convolution block CW is followed by the first group convolution block GW. The convolution kernel size of the first group convolution block GW is 3*3, the step size s=1, the number of groups group=8, the output channel is 8, and the output size of the 56*56 face image remains unchanged after passing through the first group convolution block GW, that is, the output size is 56*56.

[0046] After the first grouped convolution block GW, 15 Bottleneck bottleneck layers are cascaded; in this embodiment, each Bottleneck bottleneck layer is divided into a first depth-separable convolution DW and a residual structure Residual. Compared with the Bottleneck bottleneck layer in the original MobileFaceNet network, the number of channels of the bottleneck layer in this embodiment is greatly reduced. The first depth-separable convolution DW is composed of the third point-by-point convolution block PW, the third grouped convolution block GW, and the fourth point-by-point convolution in sequence. The residual structure Residual is composed of multiple second depth-separable convolutions DW, and on this basis, there is an additional shortcut connection. The size of the feature map in the first depth-separable convolution DW changes from large to small, and the number of channels increases from small to large. However, the size of the feature map of the residual structure Residual remains unchanged, and the number of channels changes, and the number of channels decreases from many to few and then increases to the number of channels when the input is restored.

[0047] After 15 Bottleneck layers, the first point-by-point convolution block PW, 3 second grouped convolution blocks GW, and the second point-by-point convolution block PW are cascaded in sequence.

[0048] The convolution kernel size of the first point-by-point convolution block PW is 1*1, the step size s=1, and the number of output channels is 104.

[0049] The convolution kernel size of each second group convolution block GW is 3*3, the step size s=1, and the number of output channels is 104. This embodiment splits the 7*7 convolution layer in the original network into 3 second group convolution blocks GW. However, the number of parameters and calculations of the 3 3*3 second group convolution blocks GW are significantly smaller than that of a 7*7 convolution, especially when the number of channels is large. The advantage is more obvious, and the two more activation layers can have stronger nonlinear capabilities. In addition, in order to increase the fusion between feature channels, this embodiment proposes to add a channel shuffling module SH after the first two 3*3 second group convolution blocks GW. Channel shuffling first divides the channels into N groups, transforms the dimensions of the features by group, and then transposes them. The transposed matrices are divided into N groups and then spliced ​​together. At this time, the new feature map obtained is shuffled compared to the original feature map channels. Channel recombination can promote information exchange between channels, extract richer feature information, and help improve network performance.

[0050] The convolution kernel size of the second point-by-point convolution block PW is 1*1, the step size s=1, and the number of output channels is 128. The actual size of the network is 1M. Compared with the original MobileFaceNet network, the network model size is significantly reduced while maintaining the recognition accuracy of the network.

[0051] In step S2, the parameters in the improved face recognition model are optimized by using the Arcface loss function based on dynamic boundaries.

[0052] Each face image finally output by the improved backbone feature extraction network is a 128-dimensional feature x. The 128-dimensional feature x is flattened by the fully connected layer and then sent to the Softmax classification layer to realize face feature classification. When performing classification, the Arcface loss function of the Softmax classification layer is improved. This embodiment designs an Arcface loss function based on dynamic boundaries, changes the fixed angle boundary to the parameter Δm of the dynamic angle boundary, and dynamically adjusts the parameter Δm according to the model training effect. Adding an adaptive mechanism to the Arcface loss function can increase the inter-class distance between different IDs, reduce the intra-class distance of the same ID, and avoid overfitting of the model at the same time, thereby improving the recognition effect, specifically:

[0053] Normalize the 128-dimensional feature x of each output face image sample, normalize the weight W, and calculate xW T The predicted score logits is obtained. Weight normalization enables the feature points to be mapped onto the unit hypersphere; this embodiment converts the dot product between the weight W and the sample feature x into the similarity between two vectors, and uses cosine similarity to measure the difference between the vectors, and then determines whether the sample i and the category y belong to the same id based on the corresponding threshold; the difference between the vectors is expressed as:

[0054]

[0055] Among them, x i represents the features of the i-th face image sample, W j represents the jth column of weight W, θ j W j and x i The angle between vectors.

[0056] The existing Softmax Loss formula is expressed as:

[0057]

[0058] According to formula (1), on the basis of formula (2), the feature and weight normalization and the bias restriction of 0 are added, and the Softmax Loss formula is expressed as:

[0059]

[0060] Therefore, the classification process only concerns the angle between W and x. On this basis, the angle and feature scaling are added, and the adaptive mechanism is added to the Arcface loss function, which can reduce the overfitting of the model and improve the recognition model effect. The improved Arcface loss function Softmax Loss is expressed as:

[0061]

[0062] Where L represents the Arcface loss function based on dynamic boundaries; N represents the number of samples; e represents a constant; S represents a scaling factor; represents the angle between the feature vector of the i-th sample and the weight vector of the corresponding true category; M represents the initial angle boundary; Δm represents the parameter of the dynamic angle boundary; n represents the total number of categories; y i represents the category of the i-th sample; θ j Represents the angle between the feature vector and the j-th category weight vector.

[0063] In this embodiment, the loss of each training batch is compared with the lower threshold and the upper threshold, and the angle is adaptively adjusted according to the difference between the loss and the threshold. The parameter Δm of the dynamic angle boundary is expressed as:

[0064]

[0065] Among them, L batch represents the loss of the current training batch; θ low represents the lower limit of the loss threshold; θ hightIndicates the upper limit of the loss threshold. When the loss of the current batch is less than the lower limit of the threshold, it indicates that the samples of the current batch are relatively easy, and the angle value can be appropriately increased; when the loss of the current batch is greater than the upper limit of the threshold, it indicates that the samples of the current batch are relatively difficult, and the angle value can be appropriately reduced. The value of Δm should not be adjusted too much each time, otherwise the training process may be unstable. This Arcface loss function based on dynamic boundaries can make the model more flexible to deal with samples of different difficulty levels, helping the model to improve the overall effect.

[0066] The addition of additive angle intervals in this embodiment can further enhance the compactness within the class and increase the distance between classes. The scaling factor s can make the loss converge more quickly during training. The dynamically adjusted Arcface loss function sets an adjustment range of the angle boundary Δm on this basis, and adjusts the value of Δm according to the performance of the model on the current training batch. If the average loss of the model on a batch is lower than the preset threshold, it indicates that the samples of the current batch are relatively easy. At this time, the value of Δm is appropriately increased. If the average loss is lower than the preset threshold, it indicates that the sample difficulty is relatively high, and the value of Δm can be reduced to avoid over-optimization. The value range of each adjustment of Δm should not be too large, otherwise the training process may be unstable. This Arcface loss function based on dynamic boundaries can make the final face recognition model more flexible to deal with samples of different difficulty levels, and help the final face recognition model to improve the overall effect.

[0067] In step S3, the improved face recognition model is trained by an adaptive exponential decay learning rate strategy to obtain a final face recognition model.

[0068] In the training process, this embodiment designs an adaptive exponential decay learning rate strategy, updates the learning rate according to the cycle length and the current training round, uses the exponential decay factor to adaptively control the learning rate decrease speed, keeps the model training stable, and resets the learning rate after each cycle to prevent the model from falling into the local optimum. Specifically:

[0069] Set a warmup training round warmup_epoch. In the first warmup_epoch iterations, increase the learning rate linearly from the initial learning rate warmup_init_lr to warmup_end_lr. A total of n rounds are trained. The first two rounds are warmed up, and the learning rate increases from warmup_init_lr to warmup_end_lr. After the warmup is completed, the learning rate is updated periodically, with cycle_length rounds as a cycle. The current learning rate is updated according to the cycle length and the current training round. In order to adapt to the different stages of the entire training process, the learning rate strategy adds an exponential decay factor on the basis of the cycle adjustment, so that the learning rate is not only adjusted within each cycle, but also adaptively decays during the entire training process, so that the learning rate adjustment is more refined, adapting to the needs of the training acceleration stage and the saturation stage, and enabling the model to converge more effectively. The learning rate strategy formula is expressed as:

[0070]

[0071] lr=lr×(decay_rate cycle_epoch ) (6)

[0072] Among them, slow_factor is a constant decay factor, which can control the exponential decline speed and adjust the decline speed to keep the model training stable; lr is the current learning rate value obtained from the optimizer, and cycle_epoch is the current training round modulo the cycle cycle_length.

[0073] When each training cycle ends, the learning rate is reset to prevent the model from falling into the local optimum, so lr can be expressed as:

[0074] lr=warmup_end_lr×e -exp_deacy_rate×epoch (7)

[0075] Among them, epoch is the current training round. Adaptively decaying the initial learning rate according to the training round can better adapt to the dynamic changes of model training; exp_decay_rate is an exponential decay function.

[0076] In this embodiment, a total of 200 rounds of training are performed, warmup_init_lr = 1*e -6, warmup_end_lr=0.01, slow_factor=30, cycle_length=50. During training, batch_size is set to 256, and the SGD optimizer is used; in this embodiment, the exponential decay function is set to 0.01 at the beginning of training, 0.001 at the 30th round of training, and 0.0001 at the 60th round of training. The exponential decay function of this embodiment gradually decreases from 0.01 to 0.0001.

[0077] This embodiment improves the network structure based on the lightweight network MobileFaceNet and designs a more lightweight network. At the same time, the Arcface loss function based on dynamic boundaries is designed for the model training effect, and the learning rate strategy of adaptive exponential decay is used to train a small face recognition model with better performance, which can solve the problem that the network model is too large and cannot be deployed on small-scale embedded devices with limited computing resources.

[0078] Embodiment 2

[0079] This embodiment provides a face recognition model training device based on MobileFaceNet, which is used to execute a face recognition model training method based on MobileFaceNet in Embodiment 1. The recognition device includes a processor, and the processor is provided with:

[0080] An improvement module is used to obtain a face recognition model based on MobileFaceNet and improve the backbone feature extraction network in the face recognition model to obtain an improved face recognition model;

[0081] An optimization module, used to optimize the parameters in the improved face recognition model through the Arcface loss function based on dynamic boundaries;

[0082] The training module is used to train the improved face recognition model through an adaptive exponential decay learning rate strategy to obtain a final face recognition model.

[0083] The processor described in this embodiment may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0084] The processor described in the present embodiment is connected with a memory for storing data or instructions. The memory may include a large capacity memory for data or instructions. For example, but not limitation, the memory may include a hard disk drive (HDD), a floppy disk drive, a solid state drive (SSD), a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive or a combination of two or more of these. In appropriate cases, the memory may include a removable or non-removable (or fixed) medium. In appropriate cases, the memory may be inside or outside the data processing device. In a specific embodiment, the memory is a non-volatile memory. In a specific embodiment, the memory includes a read-only memory (ROM) and a random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (Programmable Read-Only Memory, PROM for short), an erasable PROM (Erasable Programmable Read-Only Memory, EPROM for short), an electrically erasable PROM (Electrically Erasable Programmable Read-Only Memory, EEPROM for short), an electrically alterable ROM (Electrically Alterable Read-Only Memory, EAROM for short) or a flash memory (FLASH) or a combination of two or more of these. Under appropriate circumstances, the RAM can be a static random access memory (SRAM) or a dynamic random access memory (DRAM), wherein the DRAM can be a fast page mode dynamic random access memory (FPMDRAM), an extended data output dynamic random access memory (EDODRAM), a synchronous dynamic random access memory (SDRAM), etc.

[0085] The memory can be used to store or cache various data files that need to be processed and / or used for communication, as well as possible computer program instructions executed by the processor. The processor reads and executes the computer program instructions stored in the memory.

[0086] The processor described in this embodiment is connected to a transmission device for transmitting data with an external device. The transmission device can be used to receive or send data via a network. The above-mentioned specific examples of the network may include a wired or wireless network provided by a communication provider of the electronic device. In one example, the transmission device 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 can be a radio frequency (Radio Frequency, referred to as RF) module, which is used to communicate with the Internet wirelessly.

[0087] The processor of this embodiment is connected to an input / output device for inputting or outputting information. The input / output device is used to input or output information. In this embodiment, the input information may be a face image, and the output information may be a face image recognition result.

[0088] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A face recognition model training method based on MobileFaceNet, characterized in that: include: S1. Obtain a face recognition model based on MobileFaceNet, and improve the backbone feature extraction network in the face recognition model to obtain an improved face recognition model; S2. Optimize the parameters of the improved face recognition model through the Arcface loss function based on dynamic boundaries; S3. The improved face recognition model is trained through an adaptive exponential decay learning rate strategy to obtain the final face recognition model.

2. According to the MobileFaceNet-based face recognition model training method of claim 1, it is characterized in that: The improved backbone feature extraction network in step S1 includes: a first point-by-point convolution block and two group convolution blocks after the bottleneck layer, and a channel shuffling module is added after the two group convolution blocks.

3. According to the MobileFaceNet-based face recognition model training method of claim 1, it is characterized in that: The Arcface loss function based on dynamic boundaries in step S2: the fixed angle boundary of the Arcface loss function is changed to a dynamic angle boundary, and the parameters of the angle boundary are dynamically adjusted according to the training effect of the face recognition model.

4. According to claim 3, a face recognition model training method based on MobileFaceNet is characterized in that: The Arcface loss function based on the dynamic boundary in step S2 is expressed as: Where L represents the Arcface loss function based on dynamic boundaries; N represents the number of samples; e represents a constant; S represents a scaling factor; represents the angle between the feature vector of the i-th sample and the weight vector of the corresponding true category; M represents the initial angle boundary; Δm represents the parameter of the dynamic angle boundary; n represents the total number of categories; y i represents the category of the i-th sample; θ j Represents the angle between the feature vector and the j-th category weight vector.

5. According to claim 4, a face recognition model training method based on MobileFaceNet is characterized in that: The parameter Δm of the dynamic angle boundary is expressed as: Among them, L batch represents the loss of the current training batch; θ low represents the lower limit of the loss threshold; θ hight Indicates the upper limit of the loss threshold.

6. The face recognition model training method based on MobileFaceNet according to claim 1, characterized in that: The learning rate strategy of adaptive exponential decay in step S3 is specifically: updating the learning rate according to the training cycle length and the current training round of the face recognition model, and using the exponential decay factor to adaptively control the learning rate decrease speed to keep the face recognition model training stable.

7. A face recognition model training device based on MobileFaceNet, used to execute a face recognition model training method based on MobileFaceNet according to any one of claims 1 to 6, the training device comprising a processor, characterized in that: The processor is provided with: The improvement module is used to obtain a face recognition model based on MobileFaceNet and improve the backbone feature extraction network in the face recognition model to obtain an improved face recognition model; An optimization module, used to optimize the parameters in the improved face recognition model through the Arcface loss function based on dynamic boundaries; The training module is used to train the improved face recognition model through an adaptive exponential decay learning rate strategy to obtain a final face recognition model.

8. The face recognition model training device based on MobileFaceNet according to claim 7, characterized in that: The processor is connected to a memory for storing data or instructions.

9. The face recognition model training device based on MobileFaceNet according to claim 7, characterized in that: The processor is connected with a transmission device for transmitting data with an external device.

10. The face recognition model training device based on MobileFaceNet according to claim 7, characterized in that: The processor is connected to an input / output device for inputting or outputting information.

Images