Facial Features Localization And Face Recognition With 2D Images

Face recognition has became a very hot topic in pattern recognition field. The significance lies not only in academe but also in industrial applications. This paper involves two major researches: facial feature localization and face recognition.Facial feature localization plays a very important role in face recognition applications. The accuracy of localization has essential impact to the final recognition rate. On the basis of a wide and deep survey about related techniques, this paper researched the accurate eye localization algorithm based on eye reconstruction by principle component. The simulation results on two large face databases show that this method has satisfactory accuracy and real-time speed. Comparing to traditional methods like template match, projection function, symmetric transform, and wavelet networks, this method overcomes the challenges of lighting, pose, expression, glasses and partial occlusion. Therefore this method is suitable for real-time applications.With the fast progress of science and technology, many data with dimension of more than 10~4 come up in image processing, financing, bioinformatics, data mining etc. This presents the ’dimension disaster’ challenge to all of the traditional pattern recognition algorithms. In face recognition, the face image is a typical high-dimensional data. But only a small number of dimensions is essential to the target property. So dimension reduction should be performed to extract the most significant feature about the data. The subspace methods such as principle component analysis(PCA) and linear discriminant analysis(LDA) are the most two popular dimension reduction techniques. PCA and LDA try to seek optimal lower dimension subspaces for representing the original data under special criteria. By the objective of the criteria, these methods can be categorized into two major models: descriptive model and discriminant model. Descriptive model try to find the optimal lower dimension for data reconstruction while discriminant model try to find the optimal lower dimension for data classification. In most cases these two models are hard to merge together. We can only choose one of them. In seeking efficient face recognition algorithm this paper proposed a novel linear dimension reduction algorithm based on the samples affine space(SAS). In an artful way, this algorithm merges the optimal representation and optimal discriminant together. It projects the sample onto a subspace with the dimension of the class number. The features after dimension reduction have extreme high discriminant ability and can be applied as a classifier directly. The algorithm is instinct and effective. The simulation results on several databases show that this method outperform traditional LDA algorithm and can solve the single sample problem which makes LDA unavailable. This algorithm also shows the robustness to the change of training samples. By using the SAS method, we can make a closer explanation about the intrinsic structure of the data.