CN101923641A - An Improved Face Recognition Method - Google Patents

Abstract

An improved face recognition method belongs to the field of face recognition technology. Firstly, two or more acquired face images are processed, and multiple 2D feature points of the face area on the image are extracted, and different images are established. The corresponding relationship of the 2D feature points of the face; and then according to the constraint conditions that the points on the same spatial plane in the 3D space should be satisfied when projected on different images, it is judged whether the extracted 2D feature points of the face belong to the same spatial plane; This is to determine whether the acquired image is a planar scene, thereby identifying whether the face image comes from other two-dimensional pictures, and preventing the occurrence of deceiving the face recognition system by using other pictures and other methods. The present invention provides a method for discriminating whether an input image is from an actual face or a picture, thereby preventing the problem of being easily deceived by photos in the existing face recognition system, and improving the reliability of the face recognition system.

Description

An Improved Face Recognition Method

technical field

The invention relates to 2D image analysis and processing technology in biometric feature recognition technology, specifically a method for extracting feature points from two or more 2D images and judging whether they belong to the same spatial plane point.

Background technique

Face recognition technology is a kind of biometric recognition technology. At present, various countries attach great importance to face recognition technology, and many large companies have also launched identity authentication technology based on face recognition, which has broad application prospects in video surveillance, multimedia, process control, identity recognition and other fields. As the application of this technology gradually increases, the shortcomings of some identification technologies are also exploited. The traditional feature-based face recognition technology extracts the feature points of the face area and recognizes them according to a certain inherent relationship between special feature points. Its basis is the relationship between feature points on the image, while ignoring the spatial attributes of feature points. In the application of this kind of system, it is easy to be deceived by photos, and its reliability and security are seriously affected. The outstanding feature of the present invention is that, in the images acquired under different viewpoints according to the points on the same plane in space, the same plane constraint should be satisfied, and a stability operator, such as LMEDS, is used to reliably calculate the plane homography matrix according to the corresponding feature points , and based on this, the same-plane characteristics of feature points are discriminated.

Contents of the invention

The purpose of the present invention is to provide, by using the same-plane attributes of the feature points of multiple two-dimensional images to estimate whether the spatial regions corresponding to the feature points are on the same plane, to reliably determine whether the images are from photos rather than actual living bodies face, and then improve the reliability of face detection technology.

In order to achieve the above object, the technical solution of the present invention provides an improved face recognition method, which is characterized in that, according to the corresponding points of the spatial co-planar points on the two image planes should satisfy the plane single factor matrix constraints, the extracted The feature points of the face area are subjected to the same-plane constraint discrimination to determine whether the face image is obtained from a plane image.

Described method, it comprises the following steps:

Step 1, obtain two or more images of the same face by shooting;

Step 2, perform feature extraction processing on the acquired image to obtain a plurality of 2D feature points, and use visual matching technology to establish the corresponding relationship of feature points;

Step 3, arbitrarily extracting the corresponding feature points on the two images, and calculating the planar homography matrix; using the calculated planar homography matrix to calculate the error value of the feature points obtained in step 2;

Step 4, set an error upper limit value, if the error value of the feature point calculated in step 3 is less than or equal to the error upper limit value, it means that the feature points involved in the calculation are located in the same plane, if the error value of the feature point is greater than the set value If the upper limit of error is determined, then go to step 5;

Step 5, determine the face area in the image through the face detection method, use the feature points in the face area to recalculate the plane homography matrix, and calculate the error value of the feature points in the face area, if the error value is less than If it is equal to the set error value, it is judged that the feature points belong to the same plane, and the spatial points corresponding to the face area are planar. If the error value is greater than the set error value, it is judged that the captured image is a real face image.

The image described in step 1 is obtained by imaging the same face at different positions with a single camera, or obtained by shooting at the same time by two or more cameras, or by time-sharing acquisition of a moving face by a fixed camera

The calculation of the above-mentioned planar single factor matrix adopts the least square method estimation or LMEDS algorithm.

Compared with the existing face recognition technology, the present invention has the following advantages:

When performing face recognition, the spatial plane attributes of the feature points used for recognition are judged to determine whether these feature points belong to the same plane, so as to distinguish the picture and the actual face image, so as to avoid being deceived by the photo image and improve The safety and reliability in the use of such systems are ensured. The method proposed by the invention directly processes the two-dimensional image, does not set a priori constraints on the scene, and is more universal.

Description of drawings

Fig. 1 is a flow chart of the face recognition method proposed by the present invention.

Detailed ways

The present invention will be further described below in conjunction with drawings and embodiments.

As shown in Figure 1, an improved face recognition method includes the following steps:

1). Obtain two or more images of human faces;

2). Feature point extraction processing is performed on the acquired image to obtain multiple 2D feature points, and the corresponding relationship of feature points is established by using visual matching technology;

3). Randomly extract the corresponding feature points on the two images, and calculate the plane homography matrix H;

4). Calculate the error value of all corresponding feature points;

5). Set an error upper limit value. If the error value calculated in step 4 is less than or equal to the set value, it means that the feature points involved in the calculation belong to the same plane point. If the error value of the feature point is greater than the set value, it will be transferred to step next step;

6). Use the face detection method to determine the feature points of the face detection area, and use the feature points to calculate the plane homography matrix;

7). Calculate the error corresponding to the feature points of the face feature area;

8). Set an upper limit of error, and judge whether the error obtained in step 7 is less than or equal to the set value;

9). If yes, the feature point belongs to the same plane point, otherwise, the feature point does not belong to the same plane point, it is a real face image.

The principle of the present invention is: the same camera images the same face at two different positions, or two cameras image a face at the same time, or the camera is fixed to collect images of a moving face in time, the principle is the same , the points in the scene belonging to the same spatial plane, the coordinate positions in the two images should satisfy the constraint of the plane single factor matrix, namely:

，

为两幅图像上，同一空间点对应的图像坐标，H为平面单应矩阵。here

,

are the image coordinates corresponding to the same spatial point on the two images, and H is the plane homography matrix.

来描述，其中n是空间平面的法向矢量，d是空间平面到摄像机系圆点的距离。该空间平面参数可以用三个特征点，用线性方法计算得到。In space, three points that are not on the same straight line can uniquely determine a plane, and the space plane can be parameterized

To describe, where n is the normal vector of the space plane, d is the distance from the space plane to the camera system circle point. The space plane parameter can be obtained by using three characteristic points and calculating with a linear method.

On the image plane, define the distance from the point to the plane as:

.

.

In plane detection, 2D feature points are first extracted from the acquired image, and the correspondence between features is established.

For the feature points of interest on the image plane and their corresponding points in other images, the distance error value of each point pair is calculated according to the distance formula from the point to the plane. If the error values are not greater than a set error upper limit, it means that the spatial points corresponding to these feature points are points on the same plane, thus it can be concluded that the acquired image comes from a 2D plane face image such as a picture.

If the error value of the feature points calculated at this time is greater than the set error value, then use the face detection method to determine the feature points corresponding to the face area, and use the feature points in the face area to recalculate the plane homography matrix; then Calculate the error corresponding to the feature points in the face area; set an error upper limit, and determine whether the error of the feature points in the face area is less than or equal to the set value. If it is, it means that the feature points of the face area belong to the same plane point, and the acquired image is obtained from a picture; if otherwise, the feature point does not belong to the same plane point, then it can be concluded that the acquired image is from a real face.

For matching feature points to calculate the plane homography matrix, the LMEDS stable estimation method can be used to improve the robustness of the calculation results.

Generally speaking, the positioning of feature points in the image feature extraction process is noisy. In addition, the process of establishing the correspondence of feature points may produce factors such as wrong matching points. The present invention reduces The influence of these factors on the plane detection results makes the results more reliable.

Embodiment 1, a camera is fixed on a bracket. During the experiment, the person is located at a suitable position in front of the camera. Due to the slight movement of the human body, the camera acquires two face images at two times.

Perform feature point extraction processing on the two acquired images. Let the extracted feature points be p _i ( i =1,2,3,...,n) in the first image, and the acquired feature points in the second image Points are p ^' _i ( i =1,2,3,...,m). Use the stereo vision matching technology to establish the corresponding relationship of these feature points:

，采用最小二乘法估计得到平面单应矩阵H的值。According to the same plane constraint equation,

, using the least squares method to estimate the value of the plane homography matrix H.

Calculate the error value of each corresponding feature point:

if

为设定的误差值。则认为这些特征点对应于同一个空间平面。In the formula,

is the set error value. It is considered that these feature points correspond to the same spatial plane.

If the errors of not all feature points are less than or equal to the set value, then use the face detection technology to detect the face area in the two images, and filter out the feature points belonging to the face area from the obtained matching feature points. Only use the corresponding feature points of the face area to recalculate the plane homography matrix H _f , and calculate the error value of the corresponding feature points of the face area, and set an error upper limit ,if:

In the formula, m is the corresponding number of matching feature points belonging to the face area, and these feature points in the face area are considered to correspond to the same spatial plane. Otherwise, the acquired image is considered to be from a real human face.

In the second embodiment, the facial image is acquired by two cameras at a certain angle, and simultaneously photographing the same human face. Extract the feature points from the captured images and establish the corresponding relationship, and use the reliable LMEDS method to estimate the H matrix. Calculate the error value of each feature point and compare it with the set error upper limit value. If it is not greater than the set value, the feature point belongs to the same plane point. If it is greater than the set value, the face area is detected by face detection technology. Extract the feature points in the face area to recalculate the plane single factor matrix, calculate the error value of the feature points in the face area and compare it with the set error value, if it is less than or equal to the set value, it is considered that these points in the face area correspond to the same If it is greater than the set value, it can be determined that the acquired image is from a real human face.

The invention can realize stable and reliable plane attribute discrimination of feature points, and can prevent face recognition from being deceived by input images from 2D image sources such as photos.

Claims (3)

1. An improved face recognition method, characterized in that, comprises the following main steps: Step 1, obtain two or more images of the same face by shooting; Step 2, perform feature extraction processing on the acquired image to obtain a plurality of 2D feature points, and use visual matching technology to establish the corresponding relationship of feature points; Step 3, arbitrarily extracting the corresponding feature points on the two images, and calculating the planar homography matrix; using the calculated planar homography matrix to calculate the error value of the feature points obtained in step 2; Step 4, set an error upper limit value, if the error value of the feature point calculated in step 3 is less than or equal to the error upper limit value, it means that the feature points involved in the calculation are located in the same plane, if the error value of the feature point is greater than the set value If the upper limit of error is determined, then go to step 5; Step 5, determine the face area in the image through the face detection method, use the feature points in the face area to recalculate the plane homography matrix, and calculate the error value of the feature points in the face area, if the error value is less than If it is equal to the set error value, it is judged that the feature points belong to the same plane, and the spatial points corresponding to the face area are planar. If the error value is greater than the set error value, it is judged that the captured image is a real face image. 2. The face recognition method according to claim 1, characterized in that: the image in step 1 is obtained by imaging the same face at different positions by a single camera, or by two or more cameras at the same time Obtained by shooting, or images collected by a fixed camera on a moving face in time-sharing. 3. The face recognition method according to claim 1, characterized in that: the calculation of the plane single factor matrix adopts least square method estimation or LMEDS algorithm. the

Images