AU2019101141A4

AU2019101141A4 - Human face recognition based on Principal Component Analysis

Info

Publication number: AU2019101141A4
Application number: AU2019101141A
Authority: AU
Inventors: Jincheng Bao; Yujie HUANG; Yizhen Li; Yuepeng Li; Yongyi Xiong; Haoran Zhang
Original assignee: Huang Yujie Miss; Li Yizhen Miss
Current assignee: Huang Yujie Miss; Li Yizhen Miss
Priority date: 2019-09-30
Filing date: 2019-09-30
Publication date: 2019-10-31
Anticipated expiration: 2027-09-30

Abstract

This invention is a face recognition system lies in the field of computer vision. Since the information gained during the face recognition process is usually limited and in order to avoid interference from external redundant environmental infor mation during face recognition, our invention proposes a face recognition system under limited information condition, in which a face-region location method based on a novel color space YCgCr is proposed to realize the face detection. The recognition stage includes the following steps: Firstly, the invention uses the PCA algorithm for face recognition. The main idea is to express the largest 1-D vector of pixel constructed from 2-D facial image into the compact principal components of the feature space. Secondly, the KNN algorithm used in the invention determines the face tag according to the different distance. The average of multiple tags is used to determine which training sample the test face belongs to.

Description

This invention is a face recognition system lies in the field of computer vision. Since the information gained during the face recognition process is usually limited and in order to avoid interference from external redundant environmental information during face recognition, our invention proposes a face recognition system under limited information condition, in which a face-region location method based on a novel color space YCgCr is proposed to realize the face detection. The recognition stage includes the following steps: Firstly, the invention uses the PCA algorithm for face recognition. The main idea is to express the largest 1 -D vector of pixel constructed from 2-D facial image into the compact principal components of the feature space. Secondly, the KNN algorithm used in the invention determines the face tag according to the different distance. The average of multiple tags is used to determine which training sample the test face belongs to.

2019101141 30 Sep 2019

TITLE

Human face recognition based on Principal Component Analysis

FIELD OF THE INVENTION

This invention is in the field of digital image processing and serves as recognition of different human faces based on the method called Principal Component Analysis powered by Matlab.

BACKGROUND

Computer vision is the science of how to make machines see. It refers to the use of cameras and computers instead of human eyes to identify, track and measure objects, and further do graphics processing. This makes it easier for the human eye to observe or transmit images to an instrument for inspection. Pattern recognition is the process of processing and analyzing the information of the identified, described, classified and interpreted objects. The pattern recognition involved in the present research is mainly to identify and classify the specific patterns of pictures, photos, texts, symbols and other objects. The development and application of computer vision rely on the assistance of pattern recognition. Computer vision and pattern recognition are two important research fields in engineering science. The ultimate goal is to make computers the same visual function as human beings and the recognition ability of various objects. With the development of computer vision and pattern recognition technology, fingerprint recognition system, face recognition system, automatic driving system and other fields have been greatly developed.

With the ceaseless progress of the society and the urgent requirement of various aspects for rapid and effective automatic identification, biometrics technology has developed rapidly in recent decades. Biometrics, as an inherent attribute of human

2019101141 30 Sep 2019 beings, is the ideal basis for automatic identity authentication. Current biometric identification technologies mainly include fingerprint identification, retina identification, iris identification, gait identification, vein identification, face recognition, etc. Compared with other recognition methods, face recognition has been widely studied and applied because of its direct, friendly and convenient features. Therefore, users have no psychological barriers and they can accept it easily. In addition, we are able to further analyze the results of face recognition and obtain rich additional information about people’s gender, expression, age and so on, which expands the application prospect of face recognition.

In this invention, we use Matlab, which is a powerful tool for numerical analysis and image processing to implement the system of face recognition. The recognition system is based on Principal Component Analysis. After image preprocessing, we reshape the matrices of the images into lines and get the average face by calculating the average of the lines. Then the differences between the average face and original faces are used to build the covariance matrix. The most contributive eigenvalues and eigenvectors of the matrix are chosen to build the eigenface space. Afterwards differences are projected to the eigenface space. In the recognition process, distances of original images and projected images is calculated and the kNearest Neighbor classification algorithm is used to classify the input and recognize whether the input is among the training set.

SUMMARY

In order to meet the requirements of security check, advanced human-machine interaction and many other scenarios, this invention proposes an image recognition method for human faces based on the Primary Component Analysis(PCA). Using PCA, the redundancy of the input will be removed and facial features will be extracted for face recognition. In the process, the information in high dimension is reserved as much as possible in low dimension Consequently, this invention

2019101141 30 Sep 2019 increases the speed of processing input data remarkably.

PCA is one of the most effective mathematical methods to reduce the dimensionality of the data space. It is widely accepted in image processing and data compression. In the algorithm, linear transformation of the original data reduces the dimension of features with tiny loss of features by maximizing the variance of projected vectors. PCA is an unsupervised learning algorithm of dimension reduction, which doesn’t require human intervention and the calculation is rather small compared to other methods.

The framework of our program of recognizing faces includes: the image preprocessing part, the training part, the test part and the GUI part for testers to identify themselves.

In order to build the human face database, we collect photos of ourselves, combined with the ORL database of human facesfl]. With regard to our own photos, the program detects the face part of the whole image by a face detection method based on color-face feature and cut the face part. Then in the preprocessing part, the images are converted into grayscale and adjusted according to the histogram equalization.

In the training part, the algorithm of our invention is based on PCA. Images in the training set are reshaped into column vectors, and the average of the vectors is calculated as the eigenface. Then we calculate the covariance matrix, the product of the differences of the column vectors and the average face and its transposition. The eigenvalues of the covariance matrix indicate their contributions to express information, and the corresponding eigenvectors of some biggest eigenvalues are chosen to be the base vectors so as to reduce the dimension. The projected vectors of the differences of the column vectors and the average face constitute the eigenface space. Face images will be projected into the eigenface space.

In the test part, the k-Nearest Neighbor(kNN) algorithm is used to classify the test data. After the similar data processing in the training process, we evaluate the

2019101141 30 Sep 2019 smallest three distances of the test data and the existing training data projected in the eigenface space and get the three labels by kNN. The mode of the three labels decides the classification.

Finally, the GUI provides users with buttons to utilize the recognition system. In the interface, users is able to take photos and get the recognition result that matches in the training set. The program will take five photos consecutively and the judgement will be made 5 times individually to get a more precise recognition result. The mode of 5 recognition results will be chosen to be ultimate output. On top of that, new photos can be token in this interface and added into the training set so that the diversification of the database will be improved and consequently this invention is able to enhance its robustness and accuracy automatically.

DESCRIPTION OF DRAWING

Figure 1 shows the principle of PCA algorithm.

Figure 2 shows the data flow of how to intercept face from photo.

Figure 3 shows the data flow to culculate the projection vectors.

Figure 4 shows the figure average image.

Figure 5 shows the results of mean subtraction.

Figure 6 shows the original image in the webcams.

Figure 7 shows the workspace output.

Figure 8 shows the figure output.

DESCRIPTION OF PREFERRED EMBODIMENT

1.1 Conception of PCA Algorithm

The main idea of PCA is to use an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables. [2] These uncorrelated variables are called principle

2019101141 30 Sep 2019 components. The work of PCA is to find is to find a set of orthogonal coordinate axis from the original space sequentially. The choice of new coordinate axis is closely related to the data itself. [3] Among them, the direction with the largest variance in the original data is selected as the first new coordinate axis, the direction with the largest variance in the orthogonal plane of the first coordinate is selected as the second new coordinate axis, the direction with the largest variance in the orthogonal plane of the first and the second coordinate is selected as the third new coordinate axis. By analogy, we can obtain n such coordinate axis in the same way. Through the coordinates axis we can find that most variance are contained in the first k coordinate axis (k is a positive integer), and the latter one contains almost zero variances. Therefore, we can ignore the remaining coordinate axis and only retaining the first k coordinate axis which contains the most variances. In fact, this is equivalent to only keep the feature dimensions that contain most variances while ignoring the feature dimensions that contains almost zero variances, so as toreduce the dimension of data features.

1.2 PCA Applied in Face Recognition

The main idea of face recognition using PCA algorithm is to express the largest 1-D vector of pixel constructed from 2-D facial image into the compact principal components of the feature space. For a single facial image, we can obtain a set of coefficients by mapping the image into the feature space, which is the eigen factors of this face. If two facial images have almost the same eigen factors, these two images represent the same face. [4] The details of the procedures and algorithm will be demonstrated in the following paragraphs.

K-NEAREST NEIGHBORS (KNN)

2.1 Conception of KNN

The k-nearest neighbors algorithm is a classification algorithm used in pattern recognition. And it is one of the simplest machine learning algorithm. In KNN classification. For a given sample, the distance from it to all the training samples will be computed and the number of k closest training samples will be set as the

2019101141 30 Sep 2019 nearest neighbors of the target sample. According to the classes which the nearest neighbors belong to, the output is the class which is the most common one selected by plurality vote. The target sample will be assigned to that class to complete the classification[5].

2.2 KNN applied in image classification

For our image classification, we use Euclidean distance value to calculate the similarity of each training data to the target sample. And we set k as 3. Each pixel in a image has its own value and by applying the following formula: L(x,y)=|x_z·yi I, where x/ is the pixel value in target image,y; is the pixel value in training image We calculate the absolute deference values of the pixels and selected 3 leastvalue-images as closest neighbors. The classes of these three images are signed as class 1, class 2 and class 3 respectively. Then there exists several situations and solutions [6]:

(a) If class 1 £class2£class 3, The target sample will be classified to class1;

(b) If class 1 =class2Qclass 3, The target sample will be classified to class1;

(c) If class l£class2=class 3, The target sample will be classified to class2;

(d) If class l=class2=class 3, The target sample will be classified to class1;

FACE DETECTION

In the process of transferring training data to the database, the original image will include not only the facial part but also redundant environment information. But the training process will handle all the information in a image. So we have to import a face detection function to reshape the original image, highlighting the facial part and ignore irrelevant variables. Here we apply the First Face-region Location Based on a Novel Color Space YCgCr method to realize the face detection.

(1) Transforming RGB color space into YCbCr space by formula: gF = 0.257 * R + 0.504 * G + 0.098 * B + 16

C_r = R - Y = 0.439 * R - 9.368 * G - 0.071 * B = 128 (1) □ C_b = B - Y = -0.148 * R - 0.291 * G + 0.439 * B = 128

2019101141 30 Sep 2019

The following table explains the variables that appear in the formula, and Q,and C_r are relatively independent.

variable name meaning Y brightness

R red component

G green component

B blue component

C_r the difference between R and Y Cb the difference between B and Y (2) (2) Extracting the Feature region: the distribution of skin color points on each color component is counted. The threshold is set according to prior knowledge, and then the region of interest to be detected is segmented from the background through the threshold. The pixels in the threshold range are identified as skin color points, which are represented by 1; The pixels outside the threshold range are considered as nonskin color points, which are represented by 0.

YCrCb skin color model as shown in formula :

Hl, (if (a < Cr < β) && (γ < Cr < η)) •^skin □O, (else)

2019101141 30 Sep 2019

In our experiment, we take the range of color points as [ 140,160].

(3) Face Interception: The extracted features are projected onto the original imageto get the part of the face we need.

PROCEDURES OF FACE REGONITION BASED ON PCA ALGORITHM

4.1 Training Process

7.1.1 Data Acquisition

The image acquisition toolbox has to be installed to realize the function of taking pictures and videos in Matlab. After installing the package, we set the camera to recording a 10-frame-video in one time and store each frame as a picture in the corresponding folder. Therefore we have 10 pictures as training data for one specific target. The pictures are taken and processed by the face detection function, finally transformed into gray scale image for smaller data size and faster processing speed. The pictures are taken as gray scale image for smaller data size and faster processing speed. Meanwhile we collected some other image data by downloading on the internet, using the data base ORL-92xll2, which includes 40 persons’face and each person has 10 facial images. All the images in this database were taken as gray scale image, which is the same form with the test image data taken by ourselves. By combining the data base images online and the image taken by ourselves, we can enlarge out testing set to meet higher requirement andaccuracy.

7.1.2 Image Processing (1) There is no specific image file type required in our algorithm and we finally choose to transform all the data into format of ‘.pgm’.

(2) Transforming shape of the data: the image size taking by the camera or from the online data base will have various kinds of size. To optimize the accuracy and speed of later data processing, we reshape all the images into the size of 112*92.

(3) Dimension reduction: In computer every image is stored as a data matrix

2019101141 30 Sep 2019 (2-dimension array), we reconstruct the 2-dimension array in to a 1 dimension array. Therefore the new array is in the form of 1*10304(10204=112*92).

specific method is to move all the none-first-row number to the first row successively. The following formulas shows an example of array dimension reduction. Formula (1) is the original matrix, and Formula (2) is the reshaping matrix.

I ^{1 2 3}k .4 5 61™ iI \ ^{7 8 9} I

147258369(4) (4) Computing average face: Now all the images are reconstructed into a onedimension matrix. This matrix can be also treated as a vector. Assuming we have m images in the training data set, so there exist m vectors di,d2,. . . ,d«,

We construct a new m*n matrix including all the vectors,, where n is the total pixels in one single image(in this program we have reshape the images and set n = 10304), M is the elements each image concludes. This matrix can be treated as an all-sample matrix. The function mean() enables us to calculate the average number all the vectors and get a new average vector(yi,y2,. . . ,y«).

The average image can be displayed by reshaping the average vector in to the original image size 112*92.[7] (5) Mean subtraction

The mean subtraction is realized by subtracting the mean image vector from every image vector (di,d2,. . . ,d_w),[6] The following figures shows couple of images and ones after mean subtraction.

(6) Covariance matrix construction

After mean subtraction, we are able to obtain m new vectors d(i-_avg),d(2-_avg),.. . ,d(n-avg)· These n vectors can form a new subtraction m*n matrix A. Comput9

2019101141 30 Sep 2019 ing the following equation can obtain a covariance matrix B.

- 1 ^{B =} - di-_av_tdT__mg = —AA) (i = 1, 2.....„) (5) /-1 (7) Computing eigenvalues and eigenvectors

There several methods to compute the eigenvalue and eigenvector of a matrix. In matlab, the function eig() can be applied to calculate the eigenvalues λ_ζ and eigenvectors v_z.

(8) Selecting the best eigenvectors

The contribution rate is described as what is the proportion of the sum of selecting eigenvalues in sum of all eigenvalues. Here we set the contribution rate as 99% to select the first several eigenvalues. We remain the corresponding eigenvectors as the best eigenvectors and ignore others.

(9) Normalization

Assuming we obtain the number of p best eigenvectors. The following formula will be applied to conduct normalization of each best eigenvector.

M/ = ~^AVi(i= 1,2, ...,p) (6) λ_ζ·

2019101141 30 Sep 2019

Where μ_ζ is the normalized vectors.

(10) Eigenface

The normalized eigenvectors can be performed as a n*p matrix C, which represent the eigenface space.

(11) Projection Multiply di,d2,. . . ,d„ with matrix C to obtain the projection vectors p_z of d_z on eigenface space. The number of projection vectors dependson the number of training images.

7.1.3 Recognition (1) Reform the testing image as a l*n vector.

(2) Multiply the testing image vector with eigenface matrix C and obtain a projection vector pz.

(3) Compute the Euclidean Distance values from pz to each p_z and find out the corresponding p_z with shortest distance. The face which that specific p_zrepresents has as the most similarity with testing face. Applying the 3-nearest neighbors algorithm to realize the distance comparation.[8]

IMPROVEMENT AND DISCUSSION

After many experiments, we finally chose a precise way. We firstly transform RGB format image into YCbCr format, and then we extract the range of the original face according to the skin color range of the face. After those two steps, we transform the image into gray image and standardize the size of the image, so as to eliminate the influence of the background and exposure of the image on the result. In each experiment, we take five images in real time to from the Webcam to ensure the accuracy. First, we get the original image in the Webcams: Then we

2019101141 30 Sep 2019 get the five images after image processing:

We can figure that the result of image processing gives a fairly accurate result which can be put in use.

Then we use the images after image processing to do the recognition with the method 5.1.2.12, which will give us five result in total. Finally, we will take the result of the largest frequency as our final output.

We can see the output result matches the test object.

2019101141 30 Sep 2019

Claims

1. A human face recognition based on principal component analysis, which makes full use of PCA algorithm to realize face recognition, wherein a lot of training sets on the basis of the original data set are added, and a KNN algorithm is used in the test set to make the results accurate and reliable.

2. The human face recognition based on principal component analysis according to claim 1, wherein one PCA algorithm and KNN algorithm are introduced in order to avoid face recognition with limited information; t he KNN algorithm used in the invention determines the face tag according to the different distance; the average of multiple tags is used to determine which training sample the test face belongs to; through a single test, our invention reaches higher recognitionaccuracy than other methods.

3. The human face recognition based on principal component analysis according to claim 1, wherein during the training process, we used 42 groups of training samples, each sample contains five pictures,and each picture has a different expression; basically, the training set covering most possible conditions; s o we make surethat the results are close to optimal.

2019101141 30 Sep 2019

Figure 1

2019101141 30 Sep 2019

Figure 2

2019101141 30 Sep 2019

A total of 42 training sets, 5 in each group.

The training set pictures are all 112 x 92 pixels grayscale images.

Figure 3

2019101141 30 Sep 2019

Figure 4

2019101141 30 Sep

Figure 5

2019101141 30 Sep 2019

Figure 6

2019101141 30 Sep 2019

38 38 38 38 38

Figure 7