In recent years, independent component analysis has been widely used in the field of face recognition and other modes. Firstly, the face image is preprocessed to reduce dimension, then the ICA algorithm is used to obtain the independent base component of the face image, a separate space is used to construct a subspace, and finally the surface of the image to be recognized is used for face recognition. . Experiments were conducted on three questions: the number of training samples, the number of training individuals, and the number of independent bases that affect the recognition rate, and the results were obtained and analyzed.
introduction
Face recognition is an important part of biometric recognition. Has a very wide range of applications in the business, security and judicial industries [1]. The method of face recognition mainly implements an intelligent discrimination task through a computer. In general, face recognition of images can be defined and described as follows:
(1) Given a still or video image, perform a search, implement a judgment, and find a face;
(2) Determine the status of faces in a photo, such as gender, expression, age, etc.;
(3) Given an unknown face image, compare it with existing face images in the gallery to identify the classification.
These tasks are suitable for different working environment requirements, and some complex systems may contain all of the above. Face recognition methods are mainly based on geometric features and algebraic characteristics of the identification method, the former will identify the location of various organs, easy to understand, but poor robustness.
Based on algebraic features, the core idea is to treat face images as multidimensional (ie, the number of image pixels) vector set. These vectors are statistically identified or mapped to achieve the purpose of classification and identification. The method of distinguishing these multidimensional change maps is also the commonly used subspace method [2-3].
This paper uses the independent component analysis (ICA) method to recognize the face image, and the constructed training library has multiple expressions and postures. Its recognition has strong robustness and accuracy. The analysis and comparison of some of the processing results has wide application value.
1ICA basic principles
Independent component analysis is a more successful algorithm used in blind signal separation. It assumes that the signal is a mixture of several independent source signals. With proper projection, the second-order and higher-order statistical information can be extracted and high-order signals can be eliminated. Correlation. The basic approach is to use the ICA algorithm to solve the training set image to obtain a set of independent influencing bases. When the image to be tested is identified, it is projected in each base direction. In this way, the distance and other discrimination methods can be used to determine the category to which the test chart belongs [4].
Using ICA algorithm to recognize the face, the training sample image set X of the face can be used as a linear combination of the statistically independent base image S and the reversible mixing matrix A:
X=AS
The purpose of the ICA algorithm is to find the mixing matrix A or the discrete matrix W so that it satisfies the following equation:
Among them, I is the estimation of the independent statistical base image S.
From a statistical point of view, the non-Gaussianity of random variables is closely related to statistical independence. The strongest direction of non-Gaussianity is the projection direction of ICA. Therefore, as long as the non-Gaussian property is found, the corresponding one is the direction of the independent component to be found [5-7].
2 Algorithm Overview and Basic Processing 2.1 Image Library Selection and Image Preprocessing
The image library used in this paper is the Olivetti Research Laboratory standard face database (as shown in Figure 1). There are a total of 40 face pictures in the face database, with 10 pictures per person and a total of 400 pictures. The image format is pgm. It is a mobile bitmap file and is easy to handle. Each person's picture is from different angles, at different times and influenced by a certain light and some ornaments (such as glasses). The image size is 112×92. In many literatures, certain processing can be performed on the original image, and the image size can be reduced as much as possible while maintaining the original information of the image. It can effectively reduce the number of image pixels, which is also equivalent to a dimension reduction before subsequent processing. For example, some spectral face processing methods such as wavelet processing or the use of image segmentation to extract facial pictures can improve information validity.
For the face input data (first training sample), the gray data is arranged, that is, a matrix is ​​obtained. Each line is the end of the line representing the gray value of the training graph. The number of columns is the total number of training charts. The specific image refers to the face number and can be defined by a volume label function.
The data matrix is ​​subjected to a centralization process, that is, the mean value is processed, and then the whitening process is performed so that the whitened variable covariance matrix is ​​an identity matrix, and a covariance eigenvalue decomposition is used. The main purpose of these processes is to reduce the amount of computation [8-9].
2.2 Generation of ICA independent component basis 2.3 Face recognition
The face image data to be tested is extracted, centered and whitened, and then projected on an independent component basis generated by the method described above. Each test chart uses a cosine discriminating method to find out the most recent category. Thus categorized. The cosine formula is expressed as follows:
This is a calculation of the correlation coefficient. The larger the correlation coefficient, the higher the similarity of the two feature vectors. In this way, the vector to be measured (such as mat2) can be judged as the face category of mat1. Comparing the judgment result with the real value, the correct rate of recognition can be calculated [10].
3 experimental results and analysis
In accordance with the above method, this article tests the face, and there are several issues worth paying attention to the experimental results of the algorithm. The experiment is also based on these issues to test:
(1) The effect of training sample and test sample ratio on the correct rate of test results;
(2) The total number of training samples and test samples, the impact of the correct rate of test results;
(3) The effect of the number of independent components on the test results.
3.1 Effect of Training Sample Size on Test Results
The number of training samples obviously has a great influence on the test results. If possible, as many training sets as possible are obtained, the test results can be greatly improved. Figure 2 shows the recognition accuracy rate curves obtained for 40 people with 5, 6, 7, 8 and 9 training samples.
3.2 The effect of different data set sizes on the recognition rate
The data sets of different sizes, ie the size of the training and test population, have a certain influence on the test results. Figure 3 depicts the experimental results of this effect. It can be seen that the data set size and the accuracy rate are generally negatively correlated (not very strict), and the training and test set recognition rate of only 40 people is greatly reduced. Therefore, the reliability of the recognition result must be considered for the larger and more important face recognition process. Of course, stricter parameters are used only in the number of trainings and the number of independent components to be mentioned below.
3.3 The influence of the number of independent component components on the recognition rate
The influence of the number of independent component components on the recognition rate is twofold. On the one hand, the greater the number of independent component components, the higher the recognition accuracy, but on the other hand, too many independent component components will affect the speed of calculation. Although there are few samples in the experimental process, the speed can be tolerated, but if the training set is large and the number of independent component components is large, the recognition speed will be affected. It takes more than 3 hours for a typical personal microcomputer (2.6G dual-core) to process 400 samples of 10,000-pixel images to achieve maximum accuracy. When the demand is not high, the number of independent component components reaches 30% of the training set is sufficient to achieve better results. Figure 4 shows the results of tests with different independent component numbers. There are 6 training samples in each group and 4 testing samples. As can be seen from Figure 4, for a small number of training tests, the test success rate is relatively easy to achieve a high correct recognition rate (0.975%). It can also be said that it is easy to achieve convergence.
4 Conclusion
In this paper, the face recognition experiments were performed using independent component analysis. Through the projection of the face image into the subspace for classification and identification, the purpose of judging the face is achieved. The results show that this projection method based on independent components has a high accuracy, and a comparative experiment is conducted on some important parameter selection problems in independent component analysis in order to achieve an equilibrium in computational efficiency and accuracy [11-12 ]. The experimental results show that the number of training sets, the total number of samples and the number of independent components used have a great influence on the correctness of the calculation and present a nonlinear correlation. These results also help design face recognition software or reference when performing recognition tasks.
introduction
Face recognition is an important part of biometric recognition. Has a very wide range of applications in the business, security and judicial industries [1]. The method of face recognition mainly implements an intelligent discrimination task through a computer. In general, face recognition of images can be defined and described as follows:
(1) Given a still or video image, perform a search, implement a judgment, and find a face;
(2) Determine the status of faces in a photo, such as gender, expression, age, etc.;
(3) Given an unknown face image, compare it with existing face images in the gallery to identify the classification.
These tasks are suitable for different working environment requirements, and some complex systems may contain all of the above. Face recognition methods are mainly based on geometric features and algebraic characteristics of the identification method, the former will identify the location of various organs, easy to understand, but poor robustness.
Based on algebraic features, the core idea is to treat face images as multidimensional (ie, the number of image pixels) vector set. These vectors are statistically identified or mapped to achieve the purpose of classification and identification. The method of distinguishing these multidimensional change maps is also the commonly used subspace method [2-3].
This paper uses the independent component analysis (ICA) method to recognize the face image, and the constructed training library has multiple expressions and postures. Its recognition has strong robustness and accuracy. The analysis and comparison of some of the processing results has wide application value.
1ICA basic principles
Independent component analysis is a more successful algorithm used in blind signal separation. It assumes that the signal is a mixture of several independent source signals. With proper projection, the second-order and higher-order statistical information can be extracted and high-order signals can be eliminated. Correlation. The basic approach is to use the ICA algorithm to solve the training set image to obtain a set of independent influencing bases. When the image to be tested is identified, it is projected in each base direction. In this way, the distance and other discrimination methods can be used to determine the category to which the test chart belongs [4].
Using ICA algorithm to recognize the face, the training sample image set X of the face can be used as a linear combination of the statistically independent base image S and the reversible mixing matrix A:
X=AS
The purpose of the ICA algorithm is to find the mixing matrix A or the discrete matrix W so that it satisfies the following equation:
Among them, I is the estimation of the independent statistical base image S.
From a statistical point of view, the non-Gaussianity of random variables is closely related to statistical independence. The strongest direction of non-Gaussianity is the projection direction of ICA. Therefore, as long as the non-Gaussian property is found, the corresponding one is the direction of the independent component to be found [5-7].
2 Algorithm Overview and Basic Processing 2.1 Image Library Selection and Image Preprocessing
The image library used in this paper is the Olivetti Research Laboratory standard face database (as shown in Figure 1). There are a total of 40 face pictures in the face database, with 10 pictures per person and a total of 400 pictures. The image format is pgm. It is a mobile bitmap file and is easy to handle. Each person's picture is from different angles, at different times and influenced by a certain light and some ornaments (such as glasses). The image size is 112×92. In many literatures, certain processing can be performed on the original image, and the image size can be reduced as much as possible while maintaining the original information of the image. It can effectively reduce the number of image pixels, which is also equivalent to a dimension reduction before subsequent processing. For example, some spectral face processing methods such as wavelet processing or the use of image segmentation to extract facial pictures can improve information validity.
For the face input data (first training sample), the gray data is arranged, that is, a matrix is ​​obtained. Each line is the end of the line representing the gray value of the training graph. The number of columns is the total number of training charts. The specific image refers to the face number and can be defined by a volume label function.
The data matrix is ​​subjected to a centralization process, that is, the mean value is processed, and then the whitening process is performed so that the whitened variable covariance matrix is ​​an identity matrix, and a covariance eigenvalue decomposition is used. The main purpose of these processes is to reduce the amount of computation [8-9].
2.2 Generation of ICA independent component basis 2.3 Face recognition
The face image data to be tested is extracted, centered and whitened, and then projected on an independent component basis generated by the method described above. Each test chart uses a cosine discriminating method to find out the most recent category. Thus categorized. The cosine formula is expressed as follows:
This is a calculation of the correlation coefficient. The larger the correlation coefficient, the higher the similarity of the two feature vectors. In this way, the vector to be measured (such as mat2) can be judged as the face category of mat1. Comparing the judgment result with the real value, the correct rate of recognition can be calculated [10].
3 experimental results and analysis
In accordance with the above method, this article tests the face, and there are several issues worth paying attention to the experimental results of the algorithm. The experiment is also based on these issues to test:
(1) The effect of training sample and test sample ratio on the correct rate of test results;
(2) The total number of training samples and test samples, the impact of the correct rate of test results;
(3) The effect of the number of independent components on the test results.
3.1 Effect of Training Sample Size on Test Results
The number of training samples obviously has a great influence on the test results. If possible, as many training sets as possible are obtained, the test results can be greatly improved. Figure 2 shows the recognition accuracy rate curves obtained for 40 people with 5, 6, 7, 8 and 9 training samples.
3.2 The effect of different data set sizes on the recognition rate
The data sets of different sizes, ie the size of the training and test population, have a certain influence on the test results. Figure 3 depicts the experimental results of this effect. It can be seen that the data set size and the accuracy rate are generally negatively correlated (not very strict), and the training and test set recognition rate of only 40 people is greatly reduced. Therefore, the reliability of the recognition result must be considered for the larger and more important face recognition process. Of course, stricter parameters are used only in the number of trainings and the number of independent components to be mentioned below.
3.3 The influence of the number of independent component components on the recognition rate
The influence of the number of independent component components on the recognition rate is twofold. On the one hand, the greater the number of independent component components, the higher the recognition accuracy, but on the other hand, too many independent component components will affect the speed of calculation. Although there are few samples in the experimental process, the speed can be tolerated, but if the training set is large and the number of independent component components is large, the recognition speed will be affected. It takes more than 3 hours for a typical personal microcomputer (2.6G dual-core) to process 400 samples of 10,000-pixel images to achieve maximum accuracy. When the demand is not high, the number of independent component components reaches 30% of the training set is sufficient to achieve better results. Figure 4 shows the results of tests with different independent component numbers. There are 6 training samples in each group and 4 testing samples. As can be seen from Figure 4, for a small number of training tests, the test success rate is relatively easy to achieve a high correct recognition rate (0.975%). It can also be said that it is easy to achieve convergence.
4 Conclusion
In this paper, the face recognition experiments were performed using independent component analysis. Through the projection of the face image into the subspace for classification and identification, the purpose of judging the face is achieved. The results show that this projection method based on independent components has a high accuracy, and a comparative experiment is conducted on some important parameter selection problems in independent component analysis in order to achieve an equilibrium in computational efficiency and accuracy [11-12 ]. The experimental results show that the number of training sets, the total number of samples and the number of independent components used have a great influence on the correctness of the calculation and present a nonlinear correlation. These results also help design face recognition software or reference when performing recognition tasks.
Plywood for Packing,Bitangore Plywood,Poplar Core Commercial Plywood
MDF,Furniture,chipboard CO.,Ltd , http://www.nsedgebanding.com