This is to certify that I am responsible for the work submitted in this project, that the original work is my own, except as specified in the references and acknowledgments, and that the original work contained herein is not undertaken by unspecified sources or persons or not done. . Gabor Wavelets (GWs) (also known as Gabor filter) and Singular Value Decomposition (SVD) have been extensively studied in the field of face recognition. Both techniques are used to extract facial features from the human face image and presented in the form of feature vector.
Using SVD alongside the GWs increases the reliability of the facial recognition system. In the face verification and matching phase, the level of similarity between face images is determined by calculating the distance between the resulting face feature vectors obtained from GWs and SVD, respectively. Overall, the Gabor-SVD based facial recognition technique showed a constructive and promising result in recognizing the valid user and rejecting invalid users in the JAFFE database.
Without her guidance and advice I would not be able to successfully complete the Final Year Project. I would also like to thank my family, who supported me throughout the Final Year Project.
Background of Study
Even with these achievements, face recognition is still an active and attractive field in computer vision as current face recognition systems work well only under certain predefined conditions, but perform poorly when tested under different conditions such as illumination, position of header, orientation, closure, etc. Therefore, the aim of the ongoing research is to improve the performance of the face recognition system against different environments.
Problem Statement
Objective and Scope of Study
- Principal Component Analysis (PCA)
- Linear Discriminant Analysis (LDA)
- Local Binary Pattern (LBP)
- Singular Value Decomposition (SVD)
- Gabor Wavelets (GWs)
Converting 2-dimensional face image matrices to 1-dimensional image vectors results in high-dimensional vector space. Comparison between face images can be done by projecting the test faces onto the fishing space of the training faces [6]. The most important advantages of LBP features include its robustness to illumination and simplicity in calculation [10].
Matrices U and V consist of left and right singular vectors in its columns and diagonal of matrix S consists of singular values of matrix A in descending order. The singular values of matrix A are equivalent to the square root of the Eigenvalue of matrix ATA or AAT. SVD-based face recognition method is an algebraic feature extraction method, where its facial features are extracted and stored in the singular vectors of U and V, and its singular value is invariant to translation and rotation [12].
Gabor wavelets (GW) which are also known as Gabor filters are among the popular techniques used in face recognition. The size of the Gaussian envelope γ The ratio between the center frequency The biological significance and computational properties of GWs are the key factors for its wide use in automatic face recognition system. This will increase the size of the Gabor face representation as well as the redundant information stored in the face representation [19].
A filter bank with 5 degrees and 8 GW orientations is commonly used in face recognition application [20]. Certain orientations of GWs are shown to be more discriminating and significant compared to the rest of GWs [21]. The computation of the magnitude responses for each Gabor filter in the filter bank is performed in order to extract the face representation of the input image.
Some of the effective techniques used to reduce the dimension of feature vectors are Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and neural networks [23]. The degree of similarity between the input image feature vector and the database is calculated by using distance metric measurement, such as chi-square metric, cosine of principal angles, G-statistics, or log-likelihood statistics [24]. Setting a higher similarity threshold increases the security of the face recognition system, but at the same time increases the chance of rejecting a legitimate identity [26].
Project Methodology
In this phase, GWs and SVD are used to extract facial features from the identified face image. For GWs, only 12 GWs out of 40 filters are selected to extract facial features from the face images. As for SVD, only the first five singular values are selected and its associated principal right singular vectors are used to represent the facial feature.
The resulting Gabor and SVD face feature vectors will then be used in the face verification phase. In this phase, the distance between feature vectors of the test image and the training images is calculated using cosine of head angles. The distance between feature vectors of the test image and training images is known as similarity score.
The similarity score obtained by the comparison must meet or exceed the predefined similarity threshold of Gabor and SVD to confirm the identity.
Tools and Software
Key Milestone and Gantt Chart
- Feature Extraction Using GWs
- Feature Extraction Using SVD
- Performance of Gabor-SVD Based Face Recognition System
- Conclusion
- Recommendation and Future work
From Figure 15, certain scales and orientations of the Gabor filters have higher average similarity scores compared to others. In this case, the Gabor filters with an average similarity score above 89% will be selected as the salient Gabor filters for the rest of the experiment. The average similarity score of the 12 selected Gabor filters for all two face images is calculated in the right end column of Table 1.
From the average similarity scores of 12 Gabor filters, it can be seen that the lowest similarity score is between face image C and face image E, which is 89.51. Three images (image P, Q, R) as shown in Figures 16, 18 and 20 are selected as test pieces to test the performance of the selected Gabor filters. For face image Q, the highest average similarity score for 5 comparisons of 12 Gabor filters is 83.685%, which is below the defined threshold value.
Since the average similarity scores for different face images are quite high, a ratio value can be set to reduce the possibility of false acceptance rate. Test image similarity scores must meet or exceed the similarity threshold for both GW and SVD in order to be accepted as a legitimate user. Two face images are used as training images to improve the robustness of the system.
The performance of the system is evaluated in terms of correct acceptance rate and correct rejection rate using the JAFFE database as test subjects. Test image similarity results that exceed the Gabor and SVD similarity threshold are considered to be successful recognition of a valid user. Test image similarity scores lower than the Gabor and SVD similarity threshold are considered to successfully reject the invalid user.
The SVD similarity scores of the test image P1 with the first training images Q1, which is 82%, did not achieve the SVD similarity score of 85%. With this additional additional training image, the correct acceptance rate of the system is significantly improved. The selection of these singular vectors is based on the singular values of the face images.
Using SVD in addition to Gabor wavelets has really improved the reliability of the face recognition system. Including SVD left singular vectors in determining the SVD similarity score can further increase the reliability of the face recognition system.
