CN111626091A - Face image annotation method and device and computer readable storage medium - Google Patents

Abstract

The embodiment of the present invention relates to the field of computer machine learning, and discloses a face image labeling method, a device and a computer-readable storage medium. The face image labeling method includes: acquiring multiple face region images of an original image of a person; Perform feature extraction on the region image to obtain multiple face feature vectors used to characterize the identity of the person, wherein one face region image corresponds to one face feature vector; perform feature clustering on multiple face feature vectors to obtain multiple face feature vectors The category to which each face feature vector belongs to in , where the categories include positive and negative categories; the face region images corresponding to the face feature vectors belonging to the positive category are marked. The face image labeling method, device and computer-readable storage medium provided by the present invention can improve the labeling efficiency of images, ensure labeling accuracy, and reduce the labor cost of image labeling.

Description

Face image labeling method, device and computer-readable storage medium

technical field

Embodiments of the present invention relate to the field of computer machine learning, and in particular, to a face image labeling method, device, and computer-readable storage medium.

Background technique

In large-scale face recognition applications, to ensure high recognition accuracy and accurately identify face images of the same person under various age groups, various angles, illumination, and contrast, a large amount of data cleaning is required in the application development stage. , Labeling work, a character needs to prepare hundreds of standard face images (such as 112*112 in size). For the collection of standard face data, the existing solutions mainly use crawler tools to crawl a large number of public pictures from the Internet, and then use face detection algorithms to crop out all detected face images in batches, and then use professional data An annotation team or a data annotation crowdsourcing platform completes the image screening. Take a person crawling 200 images as an example. Assuming that there are 5 people on each image, 1000 face images with a size of 112*112 can be cropped in the detection stage. Among them, among these 1000 images, at least 800 images are invalid and need to be manually deleted by manual annotation.

The inventor found that there are at least the following problems in the prior art: manual deletion through manual annotation, high labor cost, low labeling efficiency, and the quality of labeling cannot be effectively guaranteed, which is insufficient to support the rapid deployment of large-scale face recognition applications.

SUMMARY OF THE INVENTION

The purpose of the embodiments of the present invention is to provide a face image labeling method, device, and computer-readable storage medium, which can reduce the labor cost of image labeling while improving the labeling efficiency of images and ensuring labeling accuracy.

In order to solve the above technical problems, embodiments of the present invention provide a method for labeling a face image, including:

Obtain multiple face region images of the original image of the person; perform feature extraction on the multiple face region images to obtain multiple face feature vectors used to characterize the identity of the person, wherein one face region image corresponds to one face feature vector ; Carry out feature clustering to a plurality of described facial feature vectors, obtain the category to which each facial feature vector belongs to in a plurality of described facial feature vectors, wherein the category includes a character that is used to characterize the corresponding facial feature vectors The identity is the positive category of the target person, and the person identity corresponding to the face feature vector is the negative category of the non-target person; the face region image corresponding to the face feature vector belonging to the positive category is marked.

An embodiment of the present invention also provides a face image labeling device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores data that can be used by the at least one processor. Instructions executed by the processor, where the instructions are executed by the at least one processor, so that the at least one processor can execute the above-mentioned method for labeling a face image.

Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above-mentioned method for labeling a face image is implemented.

Compared with the prior art, the embodiment of the present invention obtains a plurality of face feature vectors for characterizing the identity of a person by performing feature extraction on a plurality of the face region images, that is, for a face that is difficult to apply to calculation The region image is digitized to facilitate the smooth progress of the subsequent steps; by performing feature clustering on a plurality of the face feature vectors, the category to which each face feature vector of the plurality of the face feature vectors belongs is obtained, which can be determined according to the The clustering result determines whether the identity of the person corresponding to the face feature vector is a target person, so that the noise data and valid data in the face image can be quickly and accurately identified; finally, the person corresponding to the face feature vector belonging to the positive class is marked. The face area image is used to complete the labeling of face data images, which can improve the labeling efficiency, effectively ensure the labeling accuracy, reduce the time cost of manual labeling, and reduce labor costs, which is suitable for large-scale face recognition applications. Fast builds provide support.

In addition, before the labeling of the face region image corresponding to the face feature vector belonging to the positive class, the method further includes: deleting the face feature vector belonging to the negative class, and retrieving the face feature vector belonging to the positive class Carry out the feature clustering again, and determine whether there is a face feature vector belonging to the negative class in the face feature vector of the feature clustering again; if there is, repeat the above steps until the feature clustering is performed again. There are no face feature vectors belonging to the negative class.

In addition, the performing feature clustering on a plurality of the face feature vectors specifically includes: respectively taking each face feature vector in the N face feature vectors as a clustering center, and placing the i-th face When the feature vector is used as the cluster center, calculate the metric distance from the other N-1 face feature vectors in the N face feature vectors to the cluster center, where N is an integer greater than 1, and i is an integer less than or equal to N; determine whether there is a preset number of metric distances less than a preset threshold in the metric distance; if not, then determine that the i-th face feature vector belongs to the negative class; if there is, Then determine whether the number of metric distances less than the preset threshold is greater than the preset number, if it is greater than, then determine that the i-th face feature vector belongs to the positive class; if it is less than, then determine that the i-th face feature vector belongs to the negative class.

In addition, before taking each of the N face feature vectors as a clustering center, the method further includes: setting a sliding window size and a sliding step; Each face feature vector in the feature vector is used as a clustering center, which specifically includes: establishing a plurality of sliding windows according to the sliding window size, the sliding step size and the N face feature vectors, wherein each sliding window The number of face feature vectors in the window is equal to the size of the sliding window; each face feature vector in each of the sliding windows is sequentially used as the cluster center.

In addition, before establishing a plurality of sliding windows according to the sliding window size, the sliding step size and the N face feature vectors, the method further includes: randomizing the N face feature vectors.

In addition, the performing feature extraction on a plurality of the face region images to obtain a plurality of face feature vectors for characterizing the identity of a person specifically includes: sequentially inputting the plurality of the face region images into a preset neural network model, Obtain the face feature vector.

In addition, the preset neural network model includes a first-level neural network and a second-level neural network; the face feature vector is calculated by: inputting the face region image into the first-level neural network, Obtain an initial vector; input the initial vector into the second-level neural network, train the initial vector through the weight vector in the second-level neural network and preset feature parameters, and obtain the face feature vector, Wherein, the characteristic parameter is a constant greater than 0.

In addition, before the feature extraction is performed on the multiple face region images, the method further includes: performing face image data preprocessing on the multiple face region images to obtain face region images whose resolutions meet preset requirements; The performing feature extraction on a plurality of the face region images specifically includes: performing feature extraction on the face region images whose resolutions meet preset requirements.

Description of drawings

One or more embodiments are exemplified by the pictures in the corresponding drawings, and these exemplifications do not constitute limitations of the embodiments, and elements with the same reference numerals in the drawings are denoted as similar elements, Unless otherwise stated, the figures in the accompanying drawings do not constitute a scale limitation.

1 is a flowchart of a method for labeling a face image provided according to a first embodiment of the present invention;

Fig. 2 is the flow chart of the MTCNN face detection provided according to the first embodiment of the present invention;

3 is a flowchart of feature extraction of a face region image provided according to the first embodiment of the present invention;

4 is a schematic diagram of a face identity recognition provided according to the first embodiment of the present invention;

5 is a flowchart of a method for labeling a face image according to a second embodiment of the present invention;

6 is a flowchart of a method for labeling a face image provided according to a third embodiment of the present invention;

7 is a schematic diagram of the clustering of the K-neighbor algorithm provided according to the third embodiment of the present invention;

FIG. 8 is a schematic structural diagram of an apparatus for labeling a face image according to a fourth embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those of ordinary skill in the art can appreciate that, in the various embodiments of the present invention, many technical details are set forth for the reader to better understand the present invention. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in the present invention can be realized.

Unless clearly required by the context, words such as "including", "comprising" and the like throughout the specification and claims should be construed in an inclusive rather than an exclusive or exhaustive sense; that is, "including but not limited to" meaning.

In the description of the present disclosure, it should be understood that the terms "first", "second" and the like are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of the present disclosure, unless stated otherwise, "plurality" means two or more.

The first embodiment of the present invention relates to a face image labeling method. The specific process is shown in FIG. 1 , including:

S101: Acquire multiple face region images of the original image of the person.

Specifically, in this embodiment, a crawler tool is used to crawl original images of characters (such as actor stills, portraits, work photos, etc.), and the original images of characters include images of faces. For multiple, multiple face images in an original image of a person may belong to the same identity or may belong to different identities.

It is worth mentioning that, as shown in FIG. 2, in this embodiment, a multi-level neural network-based cascade face detection algorithm MTCNN (Multi-Task Cascaded Convolutional Networks) is used to detect the face region image from the original image of the person. For ease of understanding, the following is a detailed description of MTCNN:

The P-Net network predicts the bounding box (bounding box) of the face area in the original image of the person, crops and scales the image in the boundingbox area to a size of 24*24 and inputs it to the R-Net network to generate the corrected boundingbox. Crop and scale the image in the bounding box area generated by the R-Net network to a size of 48*48 and input it to the O-Net network to generate the corrected bounding box coordinates and facial features position coordinates, and the bounding box area contains people The probability value of the face. The main steps of MTCNN face detection are as follows:

(1) Determine whether there is a face area in the original image of the person:

The presence or absence of a face in a region is a binary classification problem, evaluated using the logistic regression loss function:

是模型训练样本的真实人脸区域概率，

p_i是模型预测的人脸概率，p_i∈[0，1]；

表示概率值

与概率值p_i的偏移程度，偏移越大，

越大。in,

is the real face region probability of the model training sample,

_pi is the face probability predicted by the model, pi _∈ [0, 1];

represents the probability value

The degree of deviation from the probability value p _i , the larger the deviation, the

bigger.

(2) Determine whether the location of the face region is accurate.

Specifically, the following formula is used to determine whether the position of the face region is accurate:

其中，

是模型训练样本的真实人脸区域坐标，

是模型预测的人脸区域坐标，

和

由对应区域的起始顶点坐标和区域的宽、高定义，使用欧式距离的平方度量真实坐标

与预测坐标

的偏移程度。

in,

are the coordinates of the real face region of the model training sample,

are the coordinates of the face region predicted by the model,

and

Defined by the starting vertex coordinates of the corresponding area and the width and height of the area, using the square of the Euclidean distance to measure the real coordinates

with predicted coordinates

degree of offset.

(3) Determine whether the coordinates of the facial features of the face are accurate.

Specifically, the following formula is used to determine whether the coordinates of the facial features are accurate:

其中，

是模型训练样本的人脸五官坐标，

是模型预测的人脸五官坐标，使用欧式距离的平方度量真实坐标

与预测坐标

的偏移程度。上述处理过程，提取出图像上的人脸图像位置以及人脸五官坐标位置。

in,

is the facial features coordinates of the model training sample,

is the facial facial features predicted by the model, using the square of the Euclidean distance to measure the real coordinates

with predicted coordinates

degree of offset. In the above processing process, the position of the face image on the image and the coordinate position of the facial features of the face are extracted.

Preferably, after the human face region is detected, in this embodiment, face correction may also be performed on the detected human face region. Specifically, face correction is also called face alignment, which is to uniformly rotate the face avatar to a horizontal position. Based on the face region detected in the preceding steps and the coordinate positions of the facial features (left eye, right eye, nose, left mouth corner, right mouth corner), affine transformation is performed on the face image, and the transformed face image is in Horizontal state, that is, the line between the two eyes remains horizontal, and the corrected image is scaled to 112*112 size.

More preferably, since the minimum size of the face image detected by MTCNN is 12 pixels, scaling the 12*12 image to 112*112 will cause severe distortion, and the distorted image is useless for face recognition accuracy. Therefore, in this embodiment, face image data preprocessing is also performed on the plurality of face region images, so as to obtain face region images whose resolutions meet preset requirements. Specifically, this embodiment can filter the image according to the image byte size 4*1024 (ie 4kb), and remove the image with low resolution (ie, remove the image with the image size below 4kb). It can be understood that this embodiment The method does not specifically limit the judging criteria of the image resolution. It can remove images with an image size of less than 4kb, or images with an image size of less than 5kb or 6kb, both of which can achieve the same technical effect. In this way, the distorted image can be removed before the feature extraction is performed on the face region image, thereby reducing the workload of the subsequent steps and further improving the working efficiency of the face image labeling method.

S102: Perform feature extraction on multiple face region images to obtain multiple face feature vectors used to represent character identities.

Specifically, in this embodiment, a face feature vector can be obtained in the following manner: inputting the face region image into the first-level neural network to obtain an initial vector; inputting the initial vector into the second-level neural network network, train the initial vector through the weight vector in the second-level neural network and the preset feature parameter, and obtain the face feature vector, wherein the feature parameter is a constant greater than 0.

For ease of understanding, the following is a detailed description of how the second-level neural network trains the initial vector:

Assuming that the initial vector is a 512-dimensional vector, as shown in Figure 3, x _i is the 512-dimensional feature vector output by the convolutional neural network, and w _j is the weight vector. By iterative training w _j and x _i , the angle θ between w _j and x _i vectors is reduced, so as to increase the cosine value cosθ to increase the product of w _j x _i vectors, so that the identity of the person represented by the weight vector w _j can be obtained. Higher prediction probability, and an additional parameter m is added in the training process, thereby increasing the algorithm's discriminative ability for different character identities.

Specifically, w _j is a randomly generated set of vector values, and the vector value set is used to determine the person identity of the face feature vector input to the ArcFace algorithm. For example, input feature vectors x ₁ and x ₂ respectively, if the two vectors are infinitely close to w ₁ in the weight vector group, it can be determined that x ₁ and x ₂ belong to the same person identity; if there is a large separation boundary, It can be determined that they belong to different personalities.

As shown in Figure 4, it is further elaborated that the facial features extracted by the ArcFace algorithm have the characteristics of high cohesion (belonging to the same identity) and large separation boundaries (belonging to different identities). The vector x represents the feature vector of a face image, w ₁ and w ₂ are the weight vectors after training of the ArcFace algorithm, the angle between the vector x and the weight w ₁ is θ ₁ , and the angle between the vector x and the weight w ₁ is θ ₂ , θ ₁ <θ ₂ . The calculation process of the probability of the identity of the person to which the feature vector x belongs is as follows:

w ₁ x=||w ₁ ||||x||cos(θ ₁ ); w ₂ x=||w ₂ ||||x||cos(θ ₂ ); ||w ₁ |||| x||cos(θ ₁ +m)>||w ₂ ||||x||cos(θ ₂ );

||w ₁ ||||x||cos(θ ₁ )>||w ₂ ||||x||cos(θ ₂ ).

Among them, w ₁ x and w ₂ x represent the probability that the feature vector x belongs to two person identities, respectively, and the blank area in the figure represents the improvement of the discriminativeness of the person identity by the additional parameter m. The algorithm relies on the characteristic that the cosine function cosine decreases monotonically in the [0, π] interval, and adds a non-negative m parameter during training to increase the separation boundary between faces with different identities. The feature vector extracted by the ArcFace algorithm has higher cohesion between the features of the same identity, and a larger separation boundary between the features of different identities.

S103: Perform feature clustering on multiple face feature vectors to obtain a category to which each face feature vector of the multiple face feature vectors belongs.

Specifically, the categories include a positive category for characterizing the person identity corresponding to the face feature vector as a target person, and a negative category for representing the person identity corresponding to the face feature vector as a non-target person. In this embodiment, a feature clustering method based on a sliding window can be used, and by adjusting the sliding step size and the window size, the average intra-class distance of the face feature vector is gradually reduced, and the accuracy of the clustering result is gradually improved.

S104: Label the face region image corresponding to the face feature vector belonging to the positive class.

Specifically, in this embodiment, a deep neural network is used to extract face feature vectors, and according to the principle that the feature vectors of the same identity face have higher similarity, a statistical learning method is used to perform feature clustering in a sliding window, and the clustering results are identified. It can quickly and accurately screen out face images belonging to a certain identity from a large number of face images, and remove other noise data.

Compared with the prior art, the embodiment of the present invention obtains a plurality of face feature vectors for characterizing the identity of a person by performing feature extraction on a plurality of the face region images, that is, for a face that is difficult to apply to calculation The region image is digitized to facilitate the smooth progress of the subsequent steps; by performing feature clustering on a plurality of the face feature vectors, the category to which each face feature vector of the plurality of the face feature vectors belongs is obtained, which can be determined according to the The clustering result determines whether the identity of the person corresponding to the face feature vector is a target person, so that the noise data and valid data in the face image can be quickly and accurately identified; finally, the person corresponding to the face feature vector belonging to the positive class is marked. The face area image is used to complete the labeling of face data images, which can improve the labeling efficiency, effectively ensure the labeling accuracy, reduce the time cost of manual labeling, and reduce labor costs, which is suitable for large-scale face recognition applications. Fast builds provide support.

The second embodiment of the present invention relates to a face image labeling method. The second embodiment is further improved on the basis of the first embodiment. The specific improvement is that in the second embodiment, the deletion of face feature vector belonging to the negative class, and repeatedly determine whether there is a face feature vector belonging to the negative class in the face feature vector, until there is no face feature vector belonging to the negative class in the finally obtained face feature vector, This can further improve the accuracy of the standard and ensure the quality of annotation.

The specific process of this embodiment is shown in Figure 5, including:

S201: Acquire multiple face region images of the original image of the person.

S202: Perform feature extraction on multiple face region images to obtain multiple face feature vectors used to represent character identities.

S203: Perform feature clustering on multiple face feature vectors to obtain a category to which each face feature vector of the multiple face feature vectors belongs.

S204: Delete the face feature vectors belonging to the negative class, and perform feature clustering again on the face feature vectors belonging to the positive class.

S205: Determine whether there is a face feature vector belonging to a negative class in the face feature vector for performing the feature clustering again, if yes, execute step S204; if not, execute step S206.

Specifically, in this embodiment, when it is judged for the first time that there is no face feature vector belonging to the negative class in the face feature vector, the feature clustering may be performed again on the face feature vector judged this time, and It is judged again whether there is a face feature vector belonging to the negative class in the face feature vector, and repeated for many times until the result of the multiple judgments is that there is no face feature vector belonging to the negative class in the face feature vector. In this way, the accuracy of the face image labeling method can be further improved.

S206: Label the face region image corresponding to the face feature vector belonging to the positive class.

Compared with the prior art, the embodiment of the present invention obtains a plurality of face feature vectors for characterizing the identity of a person by performing feature extraction on a plurality of the face region images, that is, for a face that is difficult to apply to calculation The region image is digitized to facilitate the smooth progress of the subsequent steps; by performing feature clustering on a plurality of the face feature vectors, the category to which each face feature vector of the plurality of the face feature vectors belongs is obtained, which can be determined according to the The clustering result determines whether the identity of the person corresponding to the face feature vector is a target person, so that the noise data and valid data in the face image can be quickly and accurately identified; finally, the person corresponding to the face feature vector belonging to the positive class is marked. The face area image is used to complete the labeling of face data images, which can improve the labeling efficiency, effectively ensure the labeling accuracy, reduce the time cost of manual labeling, and reduce labor costs, which is suitable for large-scale face recognition applications. Fast builds provide support.

The third embodiment of the present invention relates to a face image labeling method. This embodiment is an example of the first embodiment, and specifically describes that: in the first embodiment, feature clustering is performed on a plurality of the face feature vectors. The process of obtaining the category to which each face feature vector of the multiple face feature vectors belongs.

Specifically, as shown in FIG. 6 , in this embodiment, steps S301 to S310 are included, wherein steps S301 to S302 are respectively substantially the same as steps S101 to S102 in the first embodiment, and are not repeated here. Repeat. The main differences are as follows:

Steps S301 to S302 are performed.

S303: respectively use the i-th face feature vector in the N face feature vectors as the cluster center.

S304: Calculate the metric distance from the other N-1 face feature vectors in the N face feature vectors to the cluster center.

S305: Determine whether there is a metric distance smaller than a preset threshold in the metric distance, and if so, perform step S306; if not, determine that the ith face feature vector belongs to the negative class.

S306: Determine whether the number of metric distances less than the preset threshold is greater than the preset number, and if so, determine that the i-th face feature vector belongs to the positive class; if not, determine that the i-th face feature vector belongs to the negative class; determine whether i is less than N , when i is less than N, set i=i+1, and execute step S303; otherwise, the process ends.

It is worth mentioning that, due to the direct feature clustering of N face feature vectors, there may be situations where the face feature vectors corresponding to multiple consecutive non-target person images may affect the clustering results, resulting in the inconsistency of the face image annotation method. The accuracy is not high. Therefore, in this embodiment, a sliding window iterative clustering method based on image list randomization can also be used, and image list randomization can reduce the impact of continuously occurring noise data on the clustering accuracy. The feature clustering method based on sliding window, by adjusting the sliding step and window size, gradually reduces the average intra-class distance of the face feature vector, and gradually improves the accuracy of the clustering results.

That is to say, before taking each of the N face feature vectors as a clustering center, it also includes: setting a sliding window size and a sliding step; Each face feature vector in the face feature vector is used as a clustering center, which specifically includes: establishing a plurality of sliding windows according to the sliding window size, the sliding step size and the N face feature vectors, wherein each of the The number of face feature vectors in the sliding window is equal to the size of the sliding window; each face feature vector in each of the sliding windows is sequentially used as the cluster center.

For ease of understanding, the following describes in detail the face feature clustering based on the sliding window in this embodiment:

First, the principle of feature clustering in this embodiment is briefly introduced: the clustering process uses a sliding window to perform local clustering within the window, and within the sliding window, each face feature f _i ∈ {f ₁ , f ₂ , ...,f _k } vector is the cluster center, calculate the metric distance between other feature vectors f _j in the window and the cluster center f _i , and judge the category (positive or negative) of the cluster center f _i according to the threshold K ).

The clustering principle of the K-nearest neighbor algorithm is shown in Figure 7. In the figure, each triangle and square represent a feature vector, and the triangle and square respectively represent the category to which the feature vector belongs. Given the metric distance and the threshold K, the same as the circle The number of triangles with similar shape features is large, so the circular feature is identified as the triangle category.

The three core elements of the K-nearest neighbor algorithm adopted in this embodiment are distance metrics, K values, and classification decision rules are set as follows:

(1) Distance metric

The distance metric in this proposal adopts the L ₂ norm, that is, the Euclidean distance (Euclidean distance) to measure, and the Euclidean distance between feature vectors is expressed as follows:

(2) Selection of K value

The corresponding relationship between the threshold K and the sliding window size and sliding step size is as follows:

Table 1

(3) Classification decision rules

If a cluster center f _i ∈{f ₁ ,f ₂ ,...,f _k } is the clustering result of the feature vector, under the set sliding window size, sliding step size, and threshold K, if the clustering result matches the If the number of features close to f _i is less than the threshold K, the cluster center f _i will be marked as a negative class, that is, noise data; if the number of features similar to f _i in the clustering result is greater than the threshold K, the cluster center f _i will be marked For the positive class, that is, valid data.

Based on the above principles, the steps of clustering the facial features of the sliding window in this embodiment can be obtained as follows:

1. Randomize the list of all face images belonging to the "same identity" to reduce the impact of continuous noise images on the clustering results.

2. Calculate the sliding window according to the sliding step size slide stride and the sliding window size window size.

3. The feature clustering in the sliding window can be divided into the following sub-steps:

Step A: Each image in the vectorization window is a 512-dimensional feature vector.

Step B: Use the K-nearest neighbor algorithm to perform feature clustering based on the principle of minimum distance, cyclically calculate the metric distance between each feature vector and the rest of the feature vectors, and the calculation results are top N arranged from small to large. The clustering calculation process is as follows:

1) Arbitrarily select a feature vector as the cluster center, denoted as f _i ∈ {f ₁ , f ₂ ,..., f _k }, and let f _i belong to the P category.

2) Calculate the fi metric distance dist from the next feature vector f _j to. If dist<0.95, _{f j is} classified as P category, otherwise f _i belongs to P category, and f _j belongs to N category.

3) Take each feature vector f _i ∈{f ₁ ,f ₂ ,...,f _k } as the cluster center in turn, calculate the distances from the remaining feature vectors f _j to each center f _i , and repeat this step until the traversal All cluster centers.

4) For the clustering result of each cluster center f _i ∈ {f ₁ ,f ₂ ,...,f _k }, according to the clustering classification rules, if the size of the P category set is less than the threshold K, the cluster center _fi is judged as a negative class, otherwise, the cluster center _fi is judged as a positive class. The positive class is valid data, and the negative class is noise data.

Step C: One round of sliding window calculation represents one round of feature clustering iteration, repeating step A and step B to the end of the iteration, the iterative convergence condition is that in the clustering results of 3 rounds of iterations, the set judged to be a negative class is an empty set .

4. Select different sliding step sizes, sliding window sizes, and clustering threshold K in turn, follow steps 1, 2, and 3 to perform feature clustering, and delete all noise data identified as negative classes based on the classification decision results of step 3 , retaining valid data identified as positive classes.

The pseudo-code description of the clustering process based on sliding window is as follows:

Compared with the prior art, the embodiment of the present invention obtains a plurality of face feature vectors for characterizing the identity of a person by performing feature extraction on a plurality of the face region images, that is, for a face that is difficult to apply to calculation The region image is digitized to facilitate the smooth progress of the subsequent steps; by performing feature clustering on a plurality of the face feature vectors, the category to which each face feature vector of the plurality of the face feature vectors belongs is obtained, which can be determined according to the The clustering result determines whether the identity of the person corresponding to the face feature vector is a target person, so that the noise data and valid data in the face image can be quickly and accurately identified; finally, the person corresponding to the face feature vector belonging to the positive class is marked. The face area image is used to complete the labeling of face data images, which can improve the labeling efficiency, effectively ensure the labeling accuracy, reduce the time cost of manual labeling, and reduce labor costs, which is suitable for large-scale face recognition applications. Fast builds provide support.

The fourth embodiment of the present invention relates to a face image labeling device, as shown in FIG. 8 , including:

at least one processor 401; and, a memory 402 in communication with the at least one processor 401; wherein the memory 402 stores instructions executable by the at least one processor 401, the instructions being executed by the at least one processor 401 to cause the at least one processor The processor 401 can execute the above-mentioned face image labeling method.

The memory 402 and the processor 401 are connected by a bus, and the bus may include any number of interconnected buses and bridges, and the bus connects one or more processors 401 and various circuits of the memory 402 together. The bus may also connect together various other circuits, such as peripherals, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides the interface between the bus and the transceiver. A transceiver may be a single element or multiple elements, such as multiple receivers and transmitters, providing a means for communicating with various other devices over a transmission medium. The data processed by the processor 401 is transmitted on the wireless medium through the antenna, and further, the antenna also receives the data and transmits the data to the processor 401 .

The processor 401 is responsible for managing the bus and general processing, and may also provide various functions including timing, peripheral interface, voltage regulation, power management, and other control functions. The memory 402 may be used to store data used by the processor 401 when performing operations.

A fifth embodiment of the present invention relates to a computer-readable storage medium storing a computer program. The above method embodiments are implemented when the computer program is executed by the processor.

That is, those skilled in the art can understand that all or part of the steps in the method for implementing the above embodiments can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium and includes several instructions to make a device ( It may be a single chip microcomputer, a chip, etc.) or a processor (processor) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: U disk, removable hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes.

Those skilled in the art can understand that the above-mentioned embodiments are specific examples for realizing the present invention, and in practical applications, various changes in form and details can be made without departing from the spirit and the spirit of the present invention. scope.

Claims (10)

1. A method for labeling a face image is characterized by comprising the following steps:

acquiring a plurality of face area images of a figure original image;

performing feature extraction on the plurality of face region images to obtain a plurality of face feature vectors for representing the identity of a person, wherein one face region image corresponds to one face feature vector;

performing feature clustering on the plurality of face feature vectors to obtain a category to which each face feature vector in the plurality of face feature vectors belongs, wherein the category comprises a positive category for representing that a person identity corresponding to the face feature vector is a target person and a negative category for representing that a person identity corresponding to the face feature vector is a non-target person;

and labeling the face region image corresponding to the face feature vector belonging to the positive class.

2. The method for labeling a face image according to claim 1, before labeling the face region image corresponding to the face feature vector belonging to the positive class, further comprising:

deleting the face feature vectors belonging to the negative class, performing feature clustering again on the face feature vectors belonging to the positive class, and judging whether the face feature vectors belonging to the negative class exist in the face feature vectors subjected to feature clustering again;

if the facial feature vectors exist, the steps are repeated until the facial feature vectors of the feature clustering do not exist, and the facial feature vectors belong to the negative class.

3. The method for labeling a face image according to claim 1 or 2, wherein the performing feature clustering on the plurality of face feature vectors specifically comprises:

respectively taking each face feature vector in the N face feature vectors as a clustering center, and calculating the measurement distance from other N-1 face feature vectors in the N face feature vectors to the clustering center when the ith face feature vector is taken as the clustering center, wherein N is an integer greater than 1, and i is an integer less than or equal to N;

judging whether a measurement distance smaller than a preset threshold value exists in the measurement distances or not;

if not, judging that the ith personal face feature vector belongs to the negative class;

if yes, judging whether the number of the measurement distances smaller than a preset threshold value is larger than or equal to a preset number, and if yes, judging that the ith personal face feature vector belongs to the positive class; if not, the ith personal face feature vector is judged to belong to the negative class.

4. The method for labeling a human face image according to claim 3, wherein before said respectively using each of the N human face feature vectors as a cluster center, the method further comprises:

setting the size of a sliding window and the sliding step length;

the respectively using each face feature vector in the N face feature vectors as a clustering center specifically includes:

establishing a plurality of sliding windows according to the size of the sliding window, the sliding step length and the N face feature vectors, wherein the number of the face feature vectors in each sliding window is equal to the size of the sliding window;

and sequentially taking each face feature vector in each sliding window as the clustering center.

5. The method for labeling human face images according to claim 4, before establishing a plurality of sliding windows according to the sliding window size, the sliding step size and the N human face feature vectors, further comprising:

and performing randomization processing on the N face feature vectors.

6. The method for labeling a human face image according to claim 1, wherein the extracting features of the plurality of human face region images to obtain a plurality of human face feature vectors for representing the identity of a person specifically comprises:

and sequentially inputting the plurality of face region images into a preset neural network model to obtain the face feature vector.

7. The method for labeling a human face image according to claim 6, wherein the preset neural network model comprises a first-level neural network and a second-level neural network; the face feature vector is calculated in the following way:

inputting the face region image into the first-stage neural network to obtain an initial vector;

inputting the initial vector into the second-stage neural network, training the initial vector through a weight vector in the second-stage neural network and preset characteristic parameters to obtain the face characteristic vector, wherein the characteristic parameters are constants larger than 0.

8. The method for labeling a human face image according to claim 1, before extracting features of the plurality of human face region images, further comprising:

carrying out face image data preprocessing on the plurality of face area images to obtain face area images with resolution meeting preset requirements;

the feature extraction of the plurality of face region images specifically includes:

and carrying out feature extraction on the face region image with the resolution meeting the preset requirement.

9. A face image labeling device is characterized by comprising:

at least one processor; and the number of the first and second groups,

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of labeling a face image according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the method for labeling a human face image according to any one of claims 1 to 8.

Images