CN108830151A - Mask detection method based on gauss hybrid models - Google Patents

Abstract

The present invention relates to the technical field of computer vision, in particular to a mask detection method based on a Gaussian mixture model, comprising the following steps: establishing a Gaussian mixture model according to a human face image sample; screening key frames containing a human face from a video stream, and Extract the facial features in the frame; send the extracted facial features into the Gaussian mixture model for matching, and judge whether the face in the key frame wears a mask according to the matching result. By using the present invention, the following effects can be achieved: the Gaussian mixture model is used to classify the face library, which can effectively distinguish between real faces and masks; three key frame screening is used to remove redundant frames, reduce calculations, and improve detection efficiency.

Description

Mask Detection Method Based on Gaussian Mixture Model

technical field

The invention relates to the technical field of computer vision, in particular to a mask detection method based on a Gaussian mixture model.

Background technique

With the rapid development of e-commerce and mobile payment, obtaining user's face image is an effective means to prevent fraud. If the user wears a mask during the photoshoot, it is likely to lead to fraudulent behavior. However, existing face detection techniques cannot distinguish between real faces and masks well, thus, an effective mask detection technique is currently lacking.

Contents of the invention

In order to solve the above problems, the present invention proposes a mask detection method based on a Gaussian mixture model, which is used to detect whether a face in a video stream wears a mask.

A mask detection method based on a Gaussian mixture model, comprising the following steps: establishing a Gaussian mixture model according to a human face image sample; screening key frames containing human faces from a video stream, and extracting human face features from the key frames; The facial features extracted from the image are sent to the Gaussian mixture model for matching, and it is judged whether the face in the key frame is wearing a mask or not according to the matching result.

Preferably, the method for extracting human face features from the key frame is: processing the color image in the key frame into a grayscale image; utilizing a face detector to detect a face on the grayscale image, and finding The face area is drawn and marked with a rectangular frame; according to the marked rectangular frame, the face features are extracted through the face feature point detection algorithm in the rectangular frame.

Preferably, the Gaussian mixture model is:

Among them, K represents the number of models; π _k represents the weight, satisfying N(x; μ _k , ∑ _k ) is the kth component of the mixed model.

Preferably, the method for screening key frames containing human faces from video streams is: extracting video frames from video streams; utilizing face detectors to filter out redundant frames that do not contain human faces in video frames; Repeated redundant frames are filtered out from the video frames to obtain key frames.

Preferably, the method for filtering out repeated redundant frames from video frames containing human faces to obtain the key frame is: obtaining the feature value of each frame in the key frame, and substituting the feature value into the following formula to obtain the adjacent The similarity of facial features in two frames of

Among them, x _ij represents the j-th dimension feature value of the i-th frame, x _(i+1)j represents the j-th dimension feature value of the i+1-th frame; set the threshold T _f , when , then the i+1th frame is a redundant frame, delete the i+1th frame; otherwise, keep the i+1th frame.

Preferably, the method for judging whether the face in the key frame is wearing a mask is: sending the face features extracted in the key frame into a Gaussian mixture model for matching to obtain several probability densities, and obtain the maximum probability density among the several probability densities. value P _max ; set an empirical threshold T, if P _max > T, it is judged that the mask is not worn; otherwise, it is judged that the mask is worn.

By using the present invention, the following effects can be achieved:

1. The present invention uses the Gaussian mixture model to classify the face library, which can effectively distinguish real faces and masks;

2. Three times of key frame screening are used to remove redundant frames, reduce calculations, and improve detection efficiency.

Description of drawings

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

Fig. 1 is a schematic flow chart of the present invention.

Detailed ways

The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings, but the present invention is not limited to these embodiments.

The basic idea of the mask detection method based on the Gaussian mixture model of the present invention is to establish a Gaussian mixture model according to the face image sample; obtain the key frame containing the face from the video stream multiple times, and extract the face feature from the key frame; The face features extracted in the key frame are sent to the Gaussian mixture model for matching, and whether the face in the key frame wears a mask is judged according to the matching result, so as to realize the fast and effective identification of the real face and the mask.

Fig. 1 is a schematic flow chart of the present invention, according to Fig. 1, it can be seen that the present invention mainly includes the following steps.

Step 1. Establish a Gaussian mixture model based on face image samples.

Specifically, in this embodiment, the present invention uses the face database of the Chinese Academy of Sciences, from which 10,000 face images are selected as a training set, including different genders, ages, head postures, expressions, etc. First, the color image is processed into a grayscale image. Secondly, use the face detector to detect the face of the gray-scale image, find out the face area in a gray-scale image and mark it with a rectangular frame. Usually, in an image, the proportion of the face to the whole image is uncertain, and the face detector is used to mark it, which is conducive to the extraction of face features. Finally, according to the markers, face features are extracted in the image through a face feature point detection algorithm.

A Gaussian mixture model is established based on the extracted face features. In this example, the Gaussian mixture model is:

Among them, K represents the number of models; π _k represents the weight, satisfying N(x; μ _k , Σ _k ) is the kth component of the mixed model.

The Gaussian mixture model (GMM) refers to the estimation of the probability density distribution of the sample, and the model used in the estimation is the weighted sum of several Gaussian models. Each Gaussian model represents a class. The data in the sample are projected on several Gaussian models, and the probabilities of each class will be obtained respectively. Then we can select the class with the highest probability as the judgment result. In this embodiment, for the established Gaussian mixture model, after the clustering of these face feature data is completed, it will be divided into multiple categories, such as men into one category, women into one category, children into one category, age People with big faces are divided into one category, people with round faces are divided into one category, and so on. For each type of image, corresponding to a set of subband energy features of discrete wavelet two-layer transform, multiple sets of feature vectors are obtained, and the multiple sets of feature vectors are used as training samples, and the EM algorithm is used to train multiple sets of feature vectors corresponding to multiple sets of feature vectors. a mixed Gaussian model. The Gaussian mixture model is used to subdivide the facial feature data into multiple categories, making the final identification result more accurate.

Step 2. Screen key frames containing human faces from the video stream, and extract human face features from the key frames.

Specifically, in this embodiment, in order to reduce model calculations and improve detection efficiency, the video frames in the video stream are screened three times to obtain key frames containing only human faces.

The first screening: Extract video frames from the video stream at a frequency of n frames/s to obtain video frames, where n is generally 1. Extract video frames at a certain frequency in the video stream, and ensure that no faces appearing in the video stream are missed.

The second screening: Use the face detector to filter out redundant frames without human faces in the video frames. Since there are many scenes without human faces in the video stream, there will also be many video frames without human faces during the video frame extraction process. There is no face in this part of the video frame, so there is no need to perform mask detection on it. In this embodiment, the part of redundant frames without human faces is screened by the human face detector, thereby reducing the operation of the Gaussian mixture model and improving the efficiency of identifying real human faces and masks. Because the face detector has errors in the recognition process, there may be video frames without faces in the video frames after one screening. In the actual screening process, the face detector can be screened multiple times.

The third screening: filter out duplicate redundant frames from video frames containing human faces to obtain key frames. Specifically, obtain the eigenvalues of each frame in N key frames, and substitute the eigenvalues into the following formula to obtain the similarity of facial features in two adjacent frames

Among them, x _ij represents the j-th dimension feature value of the i-th frame, x _(i+1)j represents the j-th dimension feature value of the i+1-th frame; set the threshold T _f , when , then the i+1th frame is a redundant frame, delete the i+1th frame; otherwise, keep the i+1th frame. There are video frames of the same similar scene in the video frames after the second screening, so there is no need to perform mask detection on repeated video frames. In this embodiment, two video frames before and after are compared with 128-dimensional eigenvalues. If the obtained result is less than the set threshold, it is judged that the two video frames are too similar, and the latter frame is judged as a redundant frame. Since there is no need to operate on similar redundant frames, the efficiency of real face and mask discrimination is improved. In this embodiment, it is judged whether it is a redundant frame by comparing the feature values of two video frames before and after, but the applicant does not limit the method of judging redundant frames, and other image processing methods can also be used to judge it .

A frame is a still picture. In this embodiment, the method of extracting facial features from key frames is the same as that of extracting facial features from facial image samples, which will not be repeated here.

Step 3. Send the facial features extracted in the key frame to the Gaussian mixture model for matching, and judge whether the face in the key frame wears a mask according to the matching result.

Specifically, the face features extracted in the key frame are sent to the Gaussian mixture model, and the distance measurement with the cluster centers of each model is obtained to obtain several probability densities, and the maximum value P _max among the several probability densities is obtained; The threshold T, in this embodiment, T=96. If P _max >T, it is determined that the mask is not worn; otherwise, it is determined that the mask is worn.

Those skilled in the art to which the present invention belongs can make various modifications or supplements to the described specific embodiments or adopt similar methods to replace them, but they will not deviate from the spirit of the present invention or go beyond the definition of the appended claims range.

Claims (6)

1. a kind of mask detection method based on gauss hybrid models, which is characterized in that include the following steps：

Gauss hybrid models are established according to facial image sample；

The key frame containing face is screened from video flowing, and face characteristic is extracted from key frame；

The face characteristic extracted in key frame feeding gauss hybrid models are matched, are judged in key frame according to matching result Face whether wear mask.

2. the mask detection method according to claim 1 based on gauss hybrid models, which is characterized in that described from key The method of extraction face characteristic is in frame：

It is gray level image by the Color Image Processing in key frame；

Face datection is carried out to gray level image using human-face detector, human face region is found out in gray level image and utilizes rectangle frame It is marked；

According to the rectangle frame of label, face characteristic is extracted by facial feature points detection algorithm in rectangle frame.

3. the mask detection method according to claim 1 based on gauss hybrid models, which is characterized in that the Gauss is mixed Molding type is：

Wherein, K represents the quantity of model；π_kWeight is represented, is metN(x；μ_k,∑_k) it is mixed model K-th of component.

4. the mask detection method according to claim 1 based on gauss hybrid models, which is characterized in that affiliated from video The method of key frame of the screening containing face is in stream：

Video frame is extracted from video flowing；

The redundant frame that face is not contained in video frame is filtered out using human-face detector；

Duplicate redundant frame is filtered out from the video frame containing face, obtains key frame.

5. the mask detection method according to claim 4 based on gauss hybrid models, which is characterized in that described from containing Duplicate redundant frame is filtered out in the video frame of face, the method for obtaining key frame is：

The characteristic value of each frame in key frame is obtained, and characteristic value substitution following formula is acquired into face spy in two adjacent frames The similarity degree of sign

Wherein, x_ijRepresent the jth dimensional feature value of the i-th frame, x_(i+1)jRepresent the jth dimensional feature value of i+1 frame；

Threshold value T is set_f, whenWhen, then i+1 frame is redundant frame, deletes i+1 frame；Otherwise, retain i+1 frame.

6. the mask detection method according to claim 1 based on gauss hybrid models, which is characterized in that the judgement is closed The method whether face in key frame wears mask is：The face characteristic extracted in key frame is sent into gauss hybrid models to carry out Several probability density are obtained after matching, acquire the maximum value P in several probability density_max；Empirical value T is set, if P_max> T is then judged as and does not wear mask；Otherwise, then it is judged as and has worn mask.