Face Recognition Based On Randomized Subspace Feature

Face recognition is one of the most important area in computer vision, which has been widely used in military and commercial applications. Generally, kernel based face recognition methods, which project visual features into higher space by using kernel tricks, are able to achieve higher performance than the models learnt in the original feature space. However, the computational burden and memory cost for kernel matrix cast a big challenge in real time applications. Motivated by the above problems and based on the random subspace theory, we have proposed two different face recognition methods. Details of our works are described as follows.We have proposed a nonlinear principal components analysis subspace feature based face recognition algorithm. Tradition kernel principal component analysis based face recognition methods obtain promising performance. However, there are two main problems exist in the algorithm. First, when the number of training samples becomes large, the computational complexity of kernel principal component analysis algorithm would increase significantly, and it would difficult to apply the algorithm in real world applications where real time performance is demanded. Second, the feature generated by kernel principal components analysis is a holistic feature, which could lose local spatial information contained in facial images. To tackle the above problems, we have proposed an algorithm to reduce the computational cost of kernel matrix by using random nonlinear principal component analysis theory, and represent facial images with local triple pattern(LTP) feature, which can capture the local texture information and increase the discriminative power of the learnt model. Besides, to further improve the recognition performance, we have introduced metric learning scheme. Extensive experimental results indicate that the proposed method can achieve higher recognition rate with significantly reduced computational cost.We have proposed a random canonical correlation analysis based face recognition method. Generally, kernel based face recognition methods maintain a huge kernel matrix, whose computational and memory costs are both very high, especially when the number of training samples increases. To reduce the computational burden and memory requirement, we have introduced random canonical correlation analysis(random CCA) into face recognition. The random CCA samples two small set of training data and constructs random features from two views by using Nystrom scheme, then the final feature representation is obtained by projecting features onto canonical correlation basis vectors. The recognition model is trained with support vector machine(SVM). Experimental results on four facial image dataset indicate the effectiveness of our method.