Reseach Of Face Recognition Algorithm Based On Semi-supervised Learning Method

Face recognition, as one of the hotest research fields of pattern recognition, is based on facial features of persons’ biometrics information. Face Recognition System began in the 1960 s, it has last for 30 years with the development of computer technology and optical image technology. But the real application comimg to the stage, occored late in 1990 with the technology improved. The key to success lies in whether the face recognition system has sophisticated core algorithm and whether the recognition result has a practical recognition rate. It integrates artificial intelligence, machine learning and a variety of professional technologies. It is the latest application of biometrics. The core technology shows the conversion from the weak artificial intelligence to the strong artificial intelligence. And the key point is dimensionality reduction algorithms for feature extraction.The main research contents and innovation points of this paper is as follows:Firstly, this paper introduces the subspace methods and semi-supervised face recognition and mainly talks about self-training PCA and LDA approach which joins unlabeled picture to the training set. Through updating the training set, it takes advantage of the unlabeled samples which may conclude important information. The above method is linear methods for face may have non-linear character, so i propose semi-supervised kernel-based PCA algorithm(SKPCA). The algorithm map the data to non-linear space of high dimensionality. The core idea is to use non-linear transformation and kernel function to solve the identification problem. The benefits of this approach is that it can maintain the non-linear characters of the image to gains higher algorithm performance. The experiments on database show that the addition of semi-supervised learning recognition is better than unsupervised learning, the kernel method is better than linear method.Secondly, this paper introduces the locality preserving projections algorithm(LPP) algorithm and an improved semi-supervised LPP algorithm(SDLPP). LPP algorithm is a face recognition algorithm by maintaining the positional relationship between the local data by constructing a neighbor matrix. It makes close points can maintain the close position after projection. The proposed SDLPP algorithm is based on that combining direct linear discriminant method(DLDA) and semi-supervised learning. It is a new face recognition algorithm. The numerical experiments show that the addition of semi-supervised learning have better recognition.The third, this paper design a semi-supervised discriminant analysis algorithm SSDPA algorithm based on SDA combined with DCT frequency transformation. Through calculating DCT coefficients and then calculate the semi-supervised discriminant power DCT coefficients(SSDP) and last use the SSDP extract the feature. The results show that SSDPA can better compensate the shortcomings of LDA’s dimensionality restriction, and obtain better recognition performance.