October 18, 2013 at 15:02Investigation of the method of principal components and linear discriminant analysis for changing the angle and conditions of face illumination as an object of recognition
Good day to all. I am a graduate student. The theme of my dissertation is “Development of image identification methods for providing individual access in real time”.
In my first post I wrote, not from the very beginning. Here I start from the very beginning.
Recognition of a person by the image of a person stands out among biometric systems in that, firstly, special or expensive equipment is not required, and secondly, physical contact with devices is not needed. However, recognition of a person by a face image does not provide 100% identification reliability.
The peculiarity is to recognize a person by the image of a person, regardless of the change in angle and lighting conditions when shooting.
Such problems do not have an exact analytical solution. This requires the selection of key features that characterize the visual image, the determination of the relative importance of the signs by choosing their weight coefficients and taking into account the relationships between the signs. Initially, these tasks were carried out by a human expert, which took a lot of time and did not guarantee quality. In new methods, key features are selected by automatically analyzing the training sample, but nevertheless, most of the information about the characteristics is set manually. For the automatic use of such analyzers, the sample should be large enough and cover all possible situations.
Neural network methods offer a different approach to solving the problem of pattern recognition. Weights in the neural network are not calculated by solving analytical equations, but are adjusted using various training methods. Neural networks are trained on a set of training examples. Trained NS can be successfully used to recognize a person under various conditions. T.O. the use of neural networks for the task of recognizing a person from a face image is a promising direction.
Methods for identifying a person by facial image
With all the variety of different algorithms and methods of image recognition, a typical recognition method consists of three components (Fig. 1):
1. transformation of the original image into the initial representation (may include both pre-processing and mathematical transformations, for example, the calculation of the main components);
2. highlighting key characteristics (for example, the first n principal components or coefficients of a discrete cosine transform are taken);
3. classification (modeling) mechanism: cluster model, metric, neural network, etc.
Fig. 1. A diagram of the relationship of structural elements of a typical image recognition method
The existing approaches to identifying persons using images of human faces are practically well established. It is necessary to improve existing algorithms in order to optimize the time and accuracy characteristics of the search they provide by using key features automatically extracted from the person’s image (for example, gender, the presence of a beard, glasses, face angle, etc.), and thus increase speed and accuracy of the search.
The principal component
method of principal components (Principal Component Analysis, PCA) is used to compress data without significant loss of information content. It consists in a linear orthogonal transformation of an input vector X of dimension N into an output vector Y of dimension M, where N> M.
Benefits:
- if there are variations in the set of images of faces, such as race, gender, emotions, lighting, components will appear, the magnitude of which is mainly determined by these factors. Therefore, the values of the corresponding main components can be used to determine, for example, the race or gender of a person;
- storage and retrieval of images in large databases, reconstruction of images.
The main difficulty lies in the high requirements for the conditions of shooting images. Images should be obtained in close lighting conditions, the same angle (solved by adding images in different angles to the training set) and high-quality preliminary processing should be carried out, bringing the images to standard conditions.
The method of linear discriminant analysis
Using the Linear Discriminant Analysis (LDA) method, the projection of the image space onto the feature space is selected so as to minimize intra-class and maximize inter-class distance in the feature space. These methods assume that classes are linearly separable.
Advantages:
- High recognition accuracy (about 94%) was noted for a wide range of lighting conditions, various facial expressions and the presence or absence of glasses.
The problems of the method:
- however, the questions remain unclear whether this method is applicable for searching in large databases, whether the method can work when in the training sample for some individuals there is an image in only one lighting condition;
- there was also no change in angle, and experiments with changes in lighting were carried out without changing other factors. Whether this method will work with such combinations is also unknown. As in the method of own persons, here, too, high-quality preliminary processing is needed, leading the images to standard conditions.
The main goal of this work is the development of recognition methods and the construction of special-use information retrieval systems (IPS SP) that provide automatic identification of a person’s personality in real time using his face image.
Achieving the goal set in the work dictates the need to solve a number of the following main tasks:
- development of “fast” algorithms for recognizing and highlighting the main characteristics of the image of a human face, providing high reliability of identification of the search object;
- development of an algorithm for storing and coding auxiliary information characterizing a search object that provides acceptable time – space performance indicators of the IPS SP;
- development of an algorithm for reliable identification of persons based on information stored in the IPS SP database;
- development of a prototype IPS SP that implements the above algorithms in order to verify in practice the correctness of the theoretical conclusions made in this work, issuing recommendations on its further improvement based on the results of the trial operation of IPS SP.
The task was set to compare the method of principal components and the method of linear discriminant analysis (LDA). It is necessary to verify the principal component method and the LDA method.
To conduct research, C ++ Builder programs were developed that implement the principal component method and the LDA method. Experimental studies were carried out using the ORL base, the FERET base, and our own base of 15 people. All databases contained images of various registration angles, with arbitrary facial expressions, various scales and registration conditions. The purpose of this experiment was to evaluate the effectiveness of the recognition method for a different number of K classes in the database (K = 4, 15, 40, 100, 200, and 395). Assessment of recognition efficiency for a small number of classes (K = 4, 15) showed that the minimum number of images in each class should not be less than 5, since the covariance matrices used in the PCA and LDA methods
The first and second components of the reduced signs Ĺ (vxvy) are “responsible” for the rotation and posture of the head, and the third for the expression itself. Moreover, the influence of the component in the recognition process (as a result of choosing a close image) is the higher, the lower the serial number of the component.
The PCA and LDA methods were tested for the case when the source database is composed of two or more dissimilar databases. To do this, 355 classes from the FERET database were added to 40 classes of the ORL database. It should be noted that the added images had a lower resolution, dark background, different lighting and sizes, as well as significant variations in face rotations.
Due to such a difference in the source data in the space of reduced features, new features appeared, grouped in a separate area relative to the features of the ORL base.
The research results are shown in table 1.
Table 1.
Conclusion
From the above analysis it follows that to increase the probability of recognition it is advisable to use a combination of both methods. This requires further research.
Literature
1. Golovko V.A. Neurointelligence: Theory and Applications. Book 1. Organization and training of neural networks with direct and feedback connections - Brest: BPI, 1999, - 260s.
2. Samal D.I., Starovoitov V.V. - Approaches and methods for recognizing people by photo portraits. - Minsk, ITC NASB, 1998. - 54 p.
3. Samal D.I., Starovoitov V.V. Technique of automated recognition of people by photo portraits // Digital image processing. - Minsk: ITK, 1999.-S.81-85.
4. Voronovsky G.K., Makhotilo K.V., Petrashev S.N., Sergeev S.A. - Genetic algorithms, artificial neural networks and virtual reality problems. - Kharkov: Basis, 1997.
Wait for the following articles :)
In my first post I wrote, not from the very beginning. Here I start from the very beginning.
Recognition of a person by the image of a person stands out among biometric systems in that, firstly, special or expensive equipment is not required, and secondly, physical contact with devices is not needed. However, recognition of a person by a face image does not provide 100% identification reliability.
The peculiarity is to recognize a person by the image of a person, regardless of the change in angle and lighting conditions when shooting.
Such problems do not have an exact analytical solution. This requires the selection of key features that characterize the visual image, the determination of the relative importance of the signs by choosing their weight coefficients and taking into account the relationships between the signs. Initially, these tasks were carried out by a human expert, which took a lot of time and did not guarantee quality. In new methods, key features are selected by automatically analyzing the training sample, but nevertheless, most of the information about the characteristics is set manually. For the automatic use of such analyzers, the sample should be large enough and cover all possible situations.
Neural network methods offer a different approach to solving the problem of pattern recognition. Weights in the neural network are not calculated by solving analytical equations, but are adjusted using various training methods. Neural networks are trained on a set of training examples. Trained NS can be successfully used to recognize a person under various conditions. T.O. the use of neural networks for the task of recognizing a person from a face image is a promising direction.
Methods for identifying a person by facial image
With all the variety of different algorithms and methods of image recognition, a typical recognition method consists of three components (Fig. 1):
1. transformation of the original image into the initial representation (may include both pre-processing and mathematical transformations, for example, the calculation of the main components);
2. highlighting key characteristics (for example, the first n principal components or coefficients of a discrete cosine transform are taken);
3. classification (modeling) mechanism: cluster model, metric, neural network, etc.
Fig. 1. A diagram of the relationship of structural elements of a typical image recognition method
The existing approaches to identifying persons using images of human faces are practically well established. It is necessary to improve existing algorithms in order to optimize the time and accuracy characteristics of the search they provide by using key features automatically extracted from the person’s image (for example, gender, the presence of a beard, glasses, face angle, etc.), and thus increase speed and accuracy of the search.
The principal component
method of principal components (Principal Component Analysis, PCA) is used to compress data without significant loss of information content. It consists in a linear orthogonal transformation of an input vector X of dimension N into an output vector Y of dimension M, where N> M.
Benefits:
- if there are variations in the set of images of faces, such as race, gender, emotions, lighting, components will appear, the magnitude of which is mainly determined by these factors. Therefore, the values of the corresponding main components can be used to determine, for example, the race or gender of a person;
- storage and retrieval of images in large databases, reconstruction of images.
The main difficulty lies in the high requirements for the conditions of shooting images. Images should be obtained in close lighting conditions, the same angle (solved by adding images in different angles to the training set) and high-quality preliminary processing should be carried out, bringing the images to standard conditions.
The method of linear discriminant analysis
Using the Linear Discriminant Analysis (LDA) method, the projection of the image space onto the feature space is selected so as to minimize intra-class and maximize inter-class distance in the feature space. These methods assume that classes are linearly separable.
Advantages:
- High recognition accuracy (about 94%) was noted for a wide range of lighting conditions, various facial expressions and the presence or absence of glasses.
The problems of the method:
- however, the questions remain unclear whether this method is applicable for searching in large databases, whether the method can work when in the training sample for some individuals there is an image in only one lighting condition;
- there was also no change in angle, and experiments with changes in lighting were carried out without changing other factors. Whether this method will work with such combinations is also unknown. As in the method of own persons, here, too, high-quality preliminary processing is needed, leading the images to standard conditions.
The main goal of this work is the development of recognition methods and the construction of special-use information retrieval systems (IPS SP) that provide automatic identification of a person’s personality in real time using his face image.
Achieving the goal set in the work dictates the need to solve a number of the following main tasks:
- development of “fast” algorithms for recognizing and highlighting the main characteristics of the image of a human face, providing high reliability of identification of the search object;
- development of an algorithm for storing and coding auxiliary information characterizing a search object that provides acceptable time – space performance indicators of the IPS SP;
- development of an algorithm for reliable identification of persons based on information stored in the IPS SP database;
- development of a prototype IPS SP that implements the above algorithms in order to verify in practice the correctness of the theoretical conclusions made in this work, issuing recommendations on its further improvement based on the results of the trial operation of IPS SP.
The task was set to compare the method of principal components and the method of linear discriminant analysis (LDA). It is necessary to verify the principal component method and the LDA method.
To conduct research, C ++ Builder programs were developed that implement the principal component method and the LDA method. Experimental studies were carried out using the ORL base, the FERET base, and our own base of 15 people. All databases contained images of various registration angles, with arbitrary facial expressions, various scales and registration conditions. The purpose of this experiment was to evaluate the effectiveness of the recognition method for a different number of K classes in the database (K = 4, 15, 40, 100, 200, and 395). Assessment of recognition efficiency for a small number of classes (K = 4, 15) showed that the minimum number of images in each class should not be less than 5, since the covariance matrices used in the PCA and LDA methods
The first and second components of the reduced signs Ĺ (vxvy) are “responsible” for the rotation and posture of the head, and the third for the expression itself. Moreover, the influence of the component in the recognition process (as a result of choosing a close image) is the higher, the lower the serial number of the component.
The PCA and LDA methods were tested for the case when the source database is composed of two or more dissimilar databases. To do this, 355 classes from the FERET database were added to 40 classes of the ORL database. It should be noted that the added images had a lower resolution, dark background, different lighting and sizes, as well as significant variations in face rotations.
Due to such a difference in the source data in the space of reduced features, new features appeared, grouped in a separate area relative to the features of the ORL base.
The research results are shown in table 1.
Table 1.
| Number of Recognizable People | The number of images of each person | Number of images used for training | Error of the second kind (FRR) when using the correlation coefficient | |
| PCA | Lda | |||
| 4 | 10 | 5 | 0,000 | 0,000 |
| fifteen | fifteen | 5 | 0.333 | 0.063 |
| 10 | 0.230 | 0.133 | ||
| 40 | 10 | 3 | 0.330 | 0.122 |
| 5 | 0.250 | 0.155 | ||
| 7 | 0.184 | 0.033 | ||
| 100 | 20 | 5 | 0.197 | 0.056 |
| 7 | 0.176 | 0.104 | ||
| 200 | 20 | 5 | 0.173 | 0.102 |
| 7 | 0.104 | 0.083 | ||
| 395 | 20 | 5 | 0.083 | 0.064 |
| 7 | 0.080 | 0.046 | ||
Conclusion
From the above analysis it follows that to increase the probability of recognition it is advisable to use a combination of both methods. This requires further research.
Literature
1. Golovko V.A. Neurointelligence: Theory and Applications. Book 1. Organization and training of neural networks with direct and feedback connections - Brest: BPI, 1999, - 260s.
2. Samal D.I., Starovoitov V.V. - Approaches and methods for recognizing people by photo portraits. - Minsk, ITC NASB, 1998. - 54 p.
3. Samal D.I., Starovoitov V.V. Technique of automated recognition of people by photo portraits // Digital image processing. - Minsk: ITK, 1999.-S.81-85.
4. Voronovsky G.K., Makhotilo K.V., Petrashev S.N., Sergeev S.A. - Genetic algorithms, artificial neural networks and virtual reality problems. - Kharkov: Basis, 1997.
Wait for the following articles :)
