Facial Feature Extraction And Recognition

In recent years, as the rapid development of computer network technology and the process of information technology is accelerating, information security and public safety have shown their unprecedented importance. Biometric identification technology in its sole, highly reliable and high stability become a hot development direction of security certification. In many biometric (fingerprint, retina, iris, mask, tone stripes, palm prints), only the facial features can be used mode of non-contact, remote collection, extraction. This method does not infringe the privacy of users and easy to be accepted. Moreover its equipment is common, simple, low cost and therefore has a broad application prospects. Not only can be applied to the Public Security system, credit card validation, file management, human-computer interaction systems, but also can be used for locking systems and performance appraisal system and other applications.In order to satisfy people's demand of convenient, fast, safe and reliable security authentication system, we design and development a security authentication system based on face recognition. This system uses Samsung's 16/32-bit embedded RISC microprocessor S3C2440A, it has small size, low power consumption, fast, reliable and stable characteristics. Compared to systems based on PC, it has some advantages, such as anti-jamming and anti-invasion, anti-virus, anti-aging. This system is composed of two parts: First, person face recognition unit based on ARM, Second, RFIC card which carries the status information(such as the new generation of identity cards). This kind of construction avoided the complex process in the usual facial identification systerm.Face recognition technology is the core of security authentication system .In this paper, I have a deep research in the methods of pretreatment, feature extraction, design the framework of the system and carriy on the simulation to the face recognition algorithm using MATLAB, compared with the impact of factors on the recognition rate such as the number of eigenvector and classification of factors. Main activities are as follows:1. Face image preprocessing minimize the influences of the background, focal distance, light, noise and other factors. Raised the accuracy of the face recognition system. According to the actual needs, introduced the person face image pretreatment methods in detail, including histogram equalization, filtering, normalization of gray and geometry of normalized.Because images in ORL and Yale database are gray-scale images, so pretreatment does not include conversion from color images to gray image. Because PCA algorithm allows facial image to have the angle change in certain scope or to have the hair , simple background and so on, therefore only images in Yale database carries on the geometry normalization, the histogram equalizing and median filtering. Therefore the experiment mainly discussed the difference between illumination normalization and nonuse of the illumination normalization .2. Face feature extraction is an important part of face recognition research. Choose the appropriate feature extraction method which should both keep the tiny difference of different faces and the clustering of the same persons faces. In the paper, two class methods of face recognition, PCA and LDA, are introduced. Also, we discuss the nearest-neighbor and the minimum distance classifiers.Although the recognition rate of LDA and PCA is not very high, its computation load is smaller than the neural network law and the hidden Markovian modeling. Moreover this system will use the RFIC card to save the facial feature, as storage space is limited, high-dimensional facial image must be compressed to low-dimensional vector, PCA meet this demand. Therefore, the system uses PCA and LDA as feature extraction methods. As for the problem of low recognition rate will be solved by reasonable pretreatment methods and classification.3. We do five experiments using the ORL faces database and the Yale database . And analysis the test results and find the project which is good for the correct recognition rate and the recognition speed.The experimental result indicated: The recognition rate enhances unceasingly along with the eigen vector dimension increases. When the dimension achieves certain value, recognition rate change flatten out. When the characteristic dimension is small, the pretreatment to the image to be equal to the introduction disturbance and the noise, but when the number of eigen vector is reasonable, the pretreatment can enhance system's recognition rate effectively. And the nearest neighbor classification is superior to the performance of minimum distance classifier.4. I has constructed the hardware platform, gave a brief introduction of each module and Elaborated the whole system's workflow. At last this article discussed the application programming interface of MATLAB and C Programming Language , converted M file to cpp file.