Study On Technology Of Face Detection And Recognition

The demand for effective automatic identity recognition is increasingly urgent in our highly inter-connected information society. Traditional personal identification methods (e.g., passwords, PIN) suffer from a number of drawbacks and are unable to satisfy the requirement. Biometrics is an automatic identification technology using individual physiological characteristics such as fingerprint, face and iris or behavioral characteristics such as signature, speech and gait. It provides a highly reliable and robust approach to the identity recognition. Face recognition technology is a friendly and direct way to identify different person as compared with recognition technologies based on other body biometrics, such as voice, fingerprint and iris, and has large potential to develop in many applications such as identity judgment, automatic surveillance, man-machine interface and so on. For the above reasons, face recognition technology has received substantial attention from researchers in the fields of biometrics, pattern recognition, computer vision and artificial intelligence.In the process of face recognition, the face detection algorithm is firstly used to determine whether the inputs such as still images or video sequences contain faces and provides the exact coordinate of each face. After that, the feature extraction method is used to extract facial features which are compared with those in the database, and then the face recognition results are obtained. So a complete face recognition system includes two main parts: face detection and face recognition. In this thesis, we first present a comprehensive survey of the fundamental theories and key technologies in the fields of face detection and face recognition, and point out the emphases and difficulties of these two parts at present. The main content of this thesis is divided into the following topics: face detection in color spaces, face recognition with small size size problem, face recognition under variable illumanition conditions and 3D face recognition. Combining the theories of statistics with pattern recognition, we put forward and realize a set of methods and algorithms that are valuable for practical application.Face detection is a key technology in the field of face information processing. Human face has a nature structure with highly complicated variations in detail, which bring great challenges to the performance of detection algorithm. These variations lie in pose, facial expression, partial occlusion, illumination variation, rotation and complex background. There is no algorithm that can handle all these variations without any kind of limitation at present. In this thesis, we focus on face detection in color space.Face detection based on complexion informationIn this dissertation, we present a multistage face detection method from coarse to fine in order to find faces in the complicated background. This method is discussed in details as follows. Firstly, the algorithm analyzes and compares complexion's clustering in the different color space, and then establishes a complexion model based on the color space of YCgCr. Secondly, the similarity degree of each pixel on the given image to the complexion is computed and the complexion similarity image is obtained. Thirdly, aimed at the shortage of the thresholding algorithms which based on the distribution of gray level of each pixel, this dissertation generalized the character space, in which the pixels are classified, from 1D to 2D by combining the distribution information of gray levels and the special information, then improved 2D Otsu algorithm is used to obtain the original classification, finally an iterative region growing technique is applied to classify the undetermined pixels. In the following selective stage, we utilize faces' geometrical characters to roughly choose the backup areas. Finally we use Euler number and eye detection method to repeatedly choose and verify these regions, and to get the end results.Face recognition based on TLDA and TKDAAfter studied four typical variants on LDA, i.e.Fisherfaces, EFM, DLDA and NLDA, which are proposed to address the well-known small sample size problem, the dissertation revealed that the discriminant features derived from Fisherfaces, EFM, DLDA and NLDA are all incomplete. Based on the analysis, a complete discriminant analysis named TLDA is proposed in this dissertation. TLDA can be used to carry out discriminant analysis in two discriminant subspaces, that is, the null space of the within-class scatter and the nonnull space of the within-class scatter. The fact that, it can make full use of two kinds of discriminant information makes TLDA a more powerful discriminator. The proposed algorithm was tested and evaluated using the ORL face database and the UMIST face database. The experimental results are encouraging.In the dissertation, motivated by the success that SVM, kernel PCA and kernel FDA have in pattern classification tasks, the proposed TLDA was generalized to nonlinear TKDA method by integrating kernel method. Obviously, the novel TKDA method retains all merits of the TLDA method, while being able to extract the nonlinear feature. The new TKDA algorithm was tested, in terms of the simplified ability and recognition accuracy, on a more complicated subset from the extended Yale B face database. The experimental results indicate that the proposed method is not only able to simplify the distribution of the face patterns, but improves the final classification results.Illumination variation face recognition based on local normalizationAn efficient representation method insensitive to varying illumination is proposed for human face recognition. Theroretical analysis based on the human face model and the illumination model shows that the effects of varying lighting on a human face image can be modeled by a sequence of multiplicative and additive noises. Instead of computing these noises, which is very difficult for real applications, we aim to reduce or even remove their effect. In our method, a local normalization technique is applied to an image, which can effectively and efficiently elimate the effect of uneven illuminations while keeping the local stastical properties of the processed image the same as in the corresponding image under normal lighting condition. After processing, the image under varying illumination will have similar pixel values to the corresponding image that is under normal lighting condition. Then the processed images are used for face recognition. The proposed algorithm has been evaluated ased on the Yale B database and the extended Yale B database. Consistent and promising results were obtained, which show that our method can effectively eliminate the effect of uneven illumination and greatly improve the recognition results.Illumination variation face recognition based on DCT in logarithm domainAnother novel illumination normalization approach for face recognition under varying lighting conditions is proposed in this dissertation. In the proposed approach, a discrete cosine transform is employed to compensate for illumination variations in the logarithm domain. Since illumination variations mainly lie in the low-frequency band, an appropriate number of DCT coefficients are truncated to minimize variations under different lighting conditions. Experimental results on the Yale B database and the Extended Yale B database show that the proposed approach improves the performance significantly for face images with large illumination variation. Moreover, the advantage of our approach is that it does not require any modeling steps and can be easily implemented in a real-time face recognition system.Research on 3D face recognitionAfter analysed the disadvantages of 2D face recognition techniques, we give a comprehensive survey of the current 3D face recognition methods and 3D data obtaining techniques.A new algorithm that project 3D data to 2D planes is proposed in this disseration. With this algorithm, we can transform 3D face model to a series of 2D face images with different poses. The projected 2D results can be used as templates to match the input face images with different poses. Experimental results on the BJUT-3D database show that the proposed approach improves the performance significantly for face images with large pose variation.2.5D data obtaining technology by minolta vivid910 and the method that construct 3D face model with 2.5D scans are studied in detail in this disseration. A 3D face database with 10 persons is built with 2.5D scans obtained by minolta vivid910 and the experimental results on our 3D database are encouraging.Face detection and recognition is challenging and there are many problems to be solved. The challenges lie in the following three aspects. Firstly, it is difficult to extract facial features that are exclusive to identification especially while variations are taken into consideration. Secondly, although some algorithms are robust to variations, they can' t be practically applied because of the huge computational complexity. Thirdly, Face recognition technology should be fused with other technologies such as password and fingerprint recognition to design a more reliable identification system. With the development of perceptive science, psychology, computer graphics, computer vision, processing of digital image and pattern recognition, application of face recognition techniques will develop rapidly and by all means come true.