Combined Weighted Eigenface And BP-based Networks For Face Recognition

These days a fast and effective identification is claimed in more and more fields. Many kinds of human inherent biometric features become the clews of identification gradually. These features include face image, fingerprint, retina, iris, sounds, gene and so on. Compared with other features, face recognition is direct, friendly, convenient and easily accepted by users. Just because of these advantages, it has become the most prospective research field. However automatic face recognition is also one of the most challenging problems. It is involved with pattern recognition, image processing, physiology, psychology and many other domains.Different from common objects, human face is a kind of non-rigid object. That is to say, different person's faces have high similarity but one person's face may look different at different time. In addition, the face images that are captured by different ways or different conditions may have substantial difference in quality, geometry, illumination, etc. Besides there also exits the makeup and face-painting influences.First we make a retrospection of the development orientation,dominating fruit, existing problems of face recognition in the issue. Then make a comparatively detailed general summarize and introduce several classical methods which are usually used in face recognition such as Eigenface based on Principal Component Analysis, Hidden Markov models, neural networks methods, elastic graph matching method etc.In the exiting face recognition system, the methods based on transform domain have already been a hot topic in the domain of pattern recognition. In the method based on transform domain, we transform the gray matrix in high dimension space into a low dimension space, at the same time extract the features in transform domain presenting the face, then compare the features. Easily speaking, the eigenface approach projects face images onto a feature space that spans the significant variations among known face images. The coefficients obtained from projection are considered the feature presenting a face image. According to comparing the coefficients of the unknown faces with the faces in the database. We will get the most similar face.The eigenface approach makes the mean error least from the high dimension to low dimension, and also has the optimal presenting ability. In this method, we need to calculate the faces'covariance matrix. Generally the covariance matrix has a high dimension, to calculate it will have a high demand for the computer's storage ability and calculation rate. This problem becomes a bottle-neck of K-L transform. Besides that as a image statistical method, all the pixels in the face image are in the same status. That will lead to omit important message and overrate the unimportant message.In this issue we aimed at the defect of the eigenface method based on traditional principal component analyze, bring forward some amelioration to improve the identification ratio and speed: firstly, in traditional method we transform the N×N face image into N 2×1face image, in this way, when the amount of face is large, the calculation process is complicated. However in this issue suggest that we should divide the face into several pieces, then it will save a lot in calculation; Secondly, in tradition method the whole eigenface face is used as the feature space in face recognition, all the pixels in the face image are in the same status, however the bothering of position, illumination and expression will lead to a drop of the face recognition ratio. And we also find that the different feature operate differently in face recognition, for example, the eyes, mouth, nose etc. are the major features in face recognition because of their abundant texture characters. On the contrary the visor and forehead (the outline message is not including), are not as important as the eyes and so on because they are lack of variety. So we can use the information to enhance the important and correlative message and restrain the unimportant and irrespective message. In this issue we bring forward a method which distributes different weight to different part of the whole face, so the important part will get a high weight. In this way we can get a higher recognition ratio; Finally, in eigenface method we calculate the average face of all the samples, then use this average face to make a criterion of each face, while in this issue we bring forward the method to calculate the average face of each sort and then use each average to make criterion of each sort. In this way we can reserve the maximum same message in each sort and enlarge different message of different sorts.Besides the improved method mentioned above, considering neural networks method have a sound calculation ability and self-adaptive learning ability, be good at associating and integrating. So we combine the PCA and back-propagation (BP) neural networks. At first, the effective features are extracted by PCA. Then, B-P based neural networks is used to recognize facial images. Traditional BP neural network is a multi-layer feed-forward network that is based on error back-propagation algorithm and has the widest application. Although traditional BP neural network is successful, it has some disadvantages. For example, its learning convergent velocity is slow, possibility of converging to a local minimum is high and so on. Therefore, an improved BP based neural networks is proposed in this thesis and is used in facial recognition. The improved BP based neural networks adopts learning rate adaptive adjustable strategy. It can effectively shorter the learning time and accelerate convergence rate.Through the experiences on ORL and Yale face data-base we can prove that the methods mentioned above are effective in enhancing the recognition ratio.