CN102024145B - Layered recognition method and system for disguised face - Google Patents

Abstract

The face recognition technology proposed by the present invention includes two layers of classifiers. Through the study of a large number of fake face samples, a first layer classifier that can identify whether it is camouflaged or not is established, and then the two classifiers on the second layer identify Face without camouflage and camouflaged face; it combines visible light sensor and infrared sensor to obtain the visible light image and infrared image of the recognized object's face respectively, and transmits it to the computer for hierarchical recognition to determine whether the object is someone in the database or not conclusions in the database. This technology effectively solves the problem that traditional face recognition technology cannot recognize fake faces or has a low recognition rate, and has broad application prospects.

Description

A method and system for layered recognition of disguised faces

technical field

The invention relates to a biometric feature recognition technology, in particular to a method and system for layered recognition of disguised human faces.

Background technique

Face recognition is to judge whether the object is the same person as the person recorded in the database according to the face image characteristics of the recognition object. It has broad application prospects due to its advantages, such as identifying recorded terrorists or criminals in airports. However, the recognition object usually camouflages the face to cover the original features of some areas of the face, which puts high demands on the face recognition system. A fake face usually refers to a face that has been altered intentionally or unintentionally, such as wearing glasses or having a beard, makeup, excessively long hair that covers part of the face, or a scarf that covers part of the face. The existing face recognition system is very inefficient for this kind of face recognition with camouflage, and cannot meet the needs of actual use.

Contents of the invention

In view of the above situation, the present invention provides a method capable of effectively identifying fake human faces, and jointly proposes a system for applying the method.

The technical scheme adopted in the present invention can be described as:

A kind of camouflage human face layered recognition method, it is characterized in that: adopt two-layer classifier to carry out layered recognition to the recognition object face, wherein the first layer of classifier is the general camouflage feature recognition classification for distinguishing whether there is camouflage in human face The second layer classifier is an individual face feature recognition classifier for identifying the identity of the object, and the second layer classifier further includes a traditional face recognition classifier for a face without camouflage and a face recognition classifier for a face with camouflage Masquerading as a face recognition classifier, the method includes the following steps:

1. Collect the face image feature data of the identified object through the visible light sensor;

2. Carry out general camouflage feature recognition in the first layer classifier, compare the feature data of the face image collected by the visible light sensor with the general feature data of the camouflage face in the database, judge whether the target face is camouflaged, and detect the camouflage type to classify;

3. Carry out individual feature recognition on the face in the second layer classifier, and directly call the traditional face recognition classifier to recognize the face feature data collected by the visible light sensor for the face without camouflage, judge the identity of the object and output the result, Recognition ends; for faces with camouflage, go to step 4;

4. According to the classification of the camouflage type in step 2, use the infrared sensor to further collect the image features of the camouflaged part of the face of the recognition object, and combine the image feature data of the uncamouflaged part of the face collected by the visible light sensor to call the camouflaged face The recognition classifier performs recognition, judges the identity of the object and outputs the result, and the recognition ends.

As an improvement of the above technical solution, the first layer classifier uses a feed-forward neural network to realize the recognition and judgment of whether the human face is disguised and the classification of the type of camouflage.

As an improvement of the above technical solution, the face image features used in the first layer classifier include: regional horizontal projection features for distinguishing whether there are glasses and beards, and regional average gray levels for distinguishing whether there are sunglasses and beards Features, the symmetric correlation coefficient and invariant moment feature for judging whether there is partial occlusion, and the local features of eyes and mouth for judging whether there is makeup or not.

As an improvement of the above technical solution, the recognition methods adopted by the traditional face recognition classifier in the second layer classifier include principal feature analysis, neural network, wavelet analysis and support vector machine algorithm.

As an improvement of the above technical solution, the camouflaged face recognition classifier in the second layer classifier adopts a neural network integration algorithm.

The present invention also proposes a camouflage human face layered recognition system, which includes an image acquisition unit for collecting the features of the face image of the object, and stores and operates the layered recognition classifier and database to carry out the collected feature data A host computer for analyzing and identifying, wherein the image acquisition unit includes parallel visible light sensors and infrared sensors, and the classifiers running on the host include two layers of classifiers, wherein the first layer of classifiers is used to distinguish whether there is camouflage on the face Masquerade recognition classifier, which uses the facial feature data collected by the visible light sensor to compare and match the general feature data of the camouflaged face in the database, judges whether the target face is camouflaged, and classifies the camouflage type; the second layer classifier It is an individual face feature recognition classifier for identifying the identity of an object, which further includes a traditional face recognition classifier for an undisguised face and a camouflaged face recognition classifier for a camouflaged face, and for a non-disguised face directly Call the traditional face recognition classifier to identify the face image feature data collected by the visible light sensor, and call the fake face recognition classifier to identify the face image feature data collected by the infrared sensor combined with the visible light sensor for a fake face. Determine the identity of the object.

As an improvement of the above technical solution, the visible light sensor (11) is a CCD or CMOS image sensor.

The beneficial effects of the present invention are:

The technology proposed by the present invention includes two layers of classifiers. By studying a large number of fake face samples, a first layer classifier that can identify whether it is camouflaged or not is established, and then the two classifiers on the second layer identify the faces without camouflage. Human face and camouflaged face; it combines visible light sensor and infrared sensor to obtain the visible light image and infrared image of the recognized object respectively, and transmits it to the computer for layered recognition to find out whether the object is someone in the database or not in the database in conclusion. This technology effectively solves the problem that traditional face recognition technology cannot recognize fake faces or has a low recognition rate, and has broad application prospects.

Description of drawings

Fig. 1 is a logic block diagram of the camouflaged face layered recognition algorithm;

Fig. 2 is a logic block diagram of the first layer classifier;

Fig. 3 is a schematic diagram of a single-layer perceptron neural network structure;

Fig. 4 is the structural representation of neural network integration;

Figure 5 is a structural block diagram of the fake face recognition system.

Detailed ways

A layered recognition method for disguised faces, as shown in the logic block diagram of the layered recognition algorithm for disguised faces in Figure 1, adopts two layers of classifiers to perform layered recognition on the faces of the recognition objects, wherein the first layer of classifiers is used to distinguish There is a general camouflage feature recognition classifier with or without camouflage on the face, and the second layer classifier is an individual face feature recognition classifier for identifying the identity of the object, and the second layer classifier further includes traditional human face recognition classifiers for faces without camouflage. A face recognition classifier and a camouflaged face recognition classifier for a camouflaged face, the method includes the following steps:

1. Collect the face image feature data of the identified object through the visible light sensor;

2. Carry out general camouflage feature recognition in the first layer classifier, compare the feature data of the face image collected by the visible light sensor with the general feature data of the camouflage face in the database, judge whether the target face is camouflaged, and detect the camouflage type to classify;

3. Carry out individual feature recognition on the face in the second layer classifier, and directly call the traditional face recognition classifier to recognize the face feature data collected by the visible light sensor for the face without camouflage, judge the identity of the object and output the result, Recognition ends; for faces with camouflage, go to step 4;

4. According to the classification of the camouflage type in step 2, use the infrared sensor to further collect the image features of the camouflaged part of the face of the recognition object, and combine the image feature data of the uncamouflaged part of the face collected by the visible light sensor to call the camouflaged face The recognition classifier performs recognition, judges the identity of the object and outputs the result, and the recognition ends.

The first layer classifier uses the feedforward neural network to classify whether the face is camouflaged and the type of camouflage. Its structure is shown in Figure 2. Its establishment process is through machine learning on a large number of face images and camouflaged face images. Face camouflage includes wearing glasses, having beards, makeup, and partially covering the face. Establish a classifier that can distinguish whether the face is camouflaged or not, which has nothing to do with the specific person. The goal of the classifier is to distinguish whether the face is camouflaged or not. The face image features used include: regional horizontal projection features are used to distinguish whether there are glasses and beards, regional average grayscale features are used to distinguish whether there are sunglasses and beards, symmetric correlation coefficients and invariant moment features are used to judge whether there are local Occlusions, local features of eyes and mouth are used to judge the presence or absence of makeup. The specific implementation of the first layer classifier is realized through a three-layer forward neural network. In Figure 2, x is the input face image feature, and nine features (x ₁ -x ₉ ) are selected: the horizontal projection feature of the eye region, Mouth area horizontal projection feature, eye area average gray level, mouth area average gray level, symmetric correlation coefficient, moment invariant feature, main feature, eye local moment invariant feature and mouth local moment invariant feature . The connection weight w _ij and bias b _ij of the neural network are obtained through machine learning. The activation function Φ _1i of the nodes in the first layer adopts the Sigmoid function, and the activation function Φ _2j of the nodes in the second layer adopts the threshold function.

In the second layer, the classifier for the recognition of undisguised faces adopts the eigenface method of principal feature analysis, and various other traditional recognition methods can also be used, such as neural networks, wavelet analysis, support vector machines, etc.

In the second layer, the classifier for camouflage face recognition uses face features not affected by camouflage combined with interpolation to recover small-area face features affected by camouflage, and recognizes face features with larger areas of occlusion combined with infrared image features. . The visible light sensor has a high resolution and collects a large amount of image information, but it is greatly affected by camouflage. Infrared sensors are less affected by camouflage, and generate infrared images of different intensities based on the temperature difference of the face, but the resolution is relatively low and the amount of image information is small. Combining the two can better identify fake faces.

The camouflage face classifier in the second layer is implemented by neural network ensemble. Neural network integration is to combine multiple single neural networks that solve the same problem into a new whole according to the design method, which is called neural network integration. A single neural network uses a single-layer perceptron network, and its structure is shown in Figure 3. According to the different types of camouflage, the features of different face images are selected, including the image features collected by the visible light sensor and the features collected by the infrared sensor.

。神经网络集成的最后总体输出为H，集成的激发函数F为：,其中

表示各类别，

，

表示取使S 为最大的C_t值。这实际上是各神经网络输出的加权多数表决结果。 The structure of neural network integration is shown in Figure 4, where x is the input feature vector of each neural network, α _i is the connection weight of each neural network linear combination to form neural network integration, and the output of each neural network participating in neural network integration is h _i ,

. The final overall output of neural network integration is H, and the integrated activation function F is: ,in

represent each category,

,

Indicates to take the value of C _t that makes S the largest. This is actually a weighted majority vote of the outputs of each neural network.

The core of neural network integration includes two parts: neural network training and classification weights:

(1) Each neural network participating in the integration is trained in turn, and the neural network trained later uses the training results of the previously trained neural network: each training sample is given a weight, and the weight of the misclassified sample in the previous neural network training is increased. The weight of samples that are correctly classified is reduced. The training optimization goal of each neural network is to minimize the sum of misclassified sample weights, so that the subsequent neural network will pay more attention to misclassified difficult samples, thereby reducing the overall classification error;

(2) The integrated weight coefficient α _i is related to the classification ability of each neural network. The smaller the weight sum of the training samples and the smaller the neural network, the greater the integrated weight coefficient α _i .

Such neural network ensemble improves the classification accuracy through the weighted combination of multiple neural network with simple structure and low precision. More importantly, this combination can also counteract the overfitting problem of too many neural network integrations.

Neural network integration is similar to a complex neural network with one hidden layer and multiple hidden layer nodes in terms of structure, but the two are fundamentally different: the two-layer neural network is machine learning as a whole neural network, so in the structure The design is complex, the number of learning samples required for machine learning is large, the learning convergence is slow, and it is easy to fall into local optimum. Each neural network of the neural network integration is designed and trained independently, and the structure of each neural network is relatively simple, so the training is relatively easy. Then proper integration of multiple neural networks with a simple structure can greatly improve the overall performance. And it can be theoretically proved that its generalization error is only related to the structural complexity of a single neural network and has nothing to do with the number of neural networks participating in the integration, which is bounded.

的调整算法如下表： The calculation of the weight coefficient α _i of the neural network integration and the training sample weight

The adjustment algorithm is as follows:

Table 1: Multi-classifier training optimization algorithm for neural network ensemble

The basic idea is that each neural network is trained sequentially, and the training sample weights of the neural network trained later are adjusted according to the training results of the trained neural network: increase the weight of samples that are misclassified, and decrease the weight of samples that have been correctly classified. This way more attention is paid to difficult samples that are misclassified. The weight coefficients of each neural network in the final integration are related to the classification accuracy of each neural network: the weight coefficient with a higher classification accuracy is larger.

为样本i的标签，

为样本i在第t个神经网络进行训练时的权重，

为第t个神经网络的输出，为第t个神经网络对样本i的输出，

为第t个神经网络在最后神经网络集成中的权重系数，

为第t个神经网络在训练中分类错误的训练样本的权重和，

、

为第t个神经网络的内部神经元间的连接权重系数和神经元的阈值参数。

为样本权重规范化因子，以使对每个神经网络的样本权重均和为1。 Let the training set be: m training samples, is the input feature of sample i,

is the label of sample i,

is the weight of sample i when training the tth neural network,

is the output of the tth neural network, is the output of the tth neural network for sample i,

is the weight coefficient of the tth neural network in the final neural network integration,

is the sum of the weights of the training samples misclassified by the t-th neural network during training,

,

is the connection weight coefficient between internal neurons of the t-th neural network and the threshold parameter of neurons.

A normalization factor for the sample weights so that the sample weights for each neural network sum to 1.

The present invention also proposes a camouflage face recognition system on the basis of the above method, the image acquisition unit 1 for collecting the face image features of the object and carrying and running the layered recognition classifier and the database for the collected features The host computer 2 for data analysis and identification is composed of two parts, as shown in FIG. 5 . Wherein the image acquisition unit includes a parallel visible light sensor 11 and an infrared sensor 12, and the classifier running on the host computer 2 includes a two-layer classifier, wherein the first layer classifier is a camouflage recognition classifier for distinguishing whether a human face is camouflaged or not, In this layer, general camouflage feature recognition is carried out, and the facial feature data collected by the visible light sensor 11 is compared and matched with the general feature data of camouflaged faces in the database to determine whether the subject's face is camouflaged, and classify the camouflage type; The second layer of classifiers is the individual face feature recognition classifier for identifying the identity of the object, which includes the traditional face recognition classifier for non-disguised faces and the camouflaged face recognition classifier for camouflaged faces. The fake face directly calls the traditional face recognition classifier to recognize the face feature data collected by the visible light sensor 11 . For the face that is judged to have camouflage in the first layer classifier, then use infrared sensor 12 to carry out further image feature collection to the camouflaged part of the recognition object face, and combine the image of the non-camera part of the face collected by the visible light sensor Feature data, calling the fake face recognition classifier for recognition, judging the identity of the object and outputting the result.

Wherein the visible light sensor 11 may be a CCD or a CMOS image sensor.

Claims (7)

1. camouflage people face layering recognition methods; It is characterized in that: adopt the two layers of classified device that identifying object face is carried out layering identification; Wherein the ground floor sorter has undisguised general camouflage feature identification sorter for being used to distinguish people's face; Second layer sorter is the individual face characteristic discriminator device that is used for the identifying object identity; Said second layer sorter comprises that further said method comprises following steps to the recognition of face sorter that does not have camouflage people face and to the camouflage recognition of face sorter that camouflage people face is arranged:

1) passes through the facial image characteristic that visible light sensor is gathered identifying object;

2) in the ground floor sorter, carry out general camouflage feature identification; The generic features data of camouflage people face compare in facial image characteristic that the use visible light sensor is gathered and the database; Judge that the object person face has or not camouflage, and the camouflage type is classified;

3) in second layer sorter, people's face is carried out personal feature identification, directly do not call the facial image characteristic that the recognition of face sorter gathers visible light sensor and discern, judge that object identity also exports the result, end of identification for there being camouflage people face; For there being camouflage people face then to get into next step;

4) according to step 2) to pretending the classification of type; Use infrared ray sensor to have the camouflage part to carry out further characteristics of image collection to identifying object people face; And combine the image feature data of the nothing camouflage part of people's face that visible light sensor gathers; Call camouflage recognition of face sorter and discern, judge object identity and export the result, end of identification.

2. a kind of camouflage people face layering recognition methods according to claim 1 is characterized in that: said ground floor sorter adopts feedforward neural network to realize the classification that whether has the identification of camouflage to judge and pretend type to people's face.

3. a kind of camouflage people face layering recognition methods according to claim 1; It is characterized in that: the facial image characteristic that adopts in the ground floor sorter comprises: being used to distinguish has the zone level of glasses-free, beard projection properties; Be used to distinguish the zone leveling gray feature that has or not sunglasses, beard; Be used to judge the symmetrical relative coefficient and the invariant moment features that have or not partial occlusion, and be used to judge the eyes that have or not cosmetic and the local feature of mouth.

4. a kind of camouflage people face layering recognition methods according to claim 1 is characterized in that: the recognition methods that the recognition of face sorter in the said second layer sorter adopts comprises neural network, wavelet analysis and algorithm of support vector machine.

5. a kind of camouflage people face layering recognition methods according to claim 1 is characterized in that: the camouflage recognition of face sorter in the second layer sorter adopts the integrated algorithm of neural network.

6. an application rights requires the camouflage people face layering recognition system of the 1 described a kind of people's of camouflage face layering recognition methods; It is characterized in that: it comprises the image acquisition units (1) that is used for acquisition target face image characteristic; And store and move said layering discriminator device and database institute's acquisition characteristics data are analyzed the main frame (2) of identification; Wherein said image acquisition units (1) comprises parallel visible light sensor (11) and infrared sensor (12); The sorter of (2) operation comprises the two layers of classified device on main frame; Wherein the ground floor sorter has undisguised camouflage discriminator device for being used to distinguish people's face; The generic features data of camouflage people face compare coupling in facial image characteristic that its use visible light sensor (11) is gathered and the database, judge that the object person face has or not camouflage, and the camouflage type is classified; Second layer sorter is the individual face characteristic discriminator device that is used for the identifying object identity; It further comprises to the recognition of face sorter that does not have camouflage people face and to the camouflage recognition of face sorter that camouflage people face is arranged; Directly do not call the facial image characteristic that the recognition of face sorter gathered visible light sensor (11) and discern for there being camouflage people face, discern to judge object identity for there being camouflage people face then to call the facial image characteristic that camouflage recognition of face sorter combines visible light sensor (11) to gather to infrared ray sensor (12).

7. a kind of camouflage people face layering recognition system according to claim 6, it is characterized in that: said visible light sensor (11) is CCD or cmos image sensor.

Images