Research On Robust Statistical Recognition Method Of Face Image With Abnormal Disturbance

The wide-range variations of human face, due to the uncertainty of pose, the modification of make up and noise jamming, result in a highly complex distribution and deteriorate the recognition performance. Although preprocessing can reduce the impact of some noise, but the abnormal disturbance of noise unavoidably prevented the further development of face recognition. These years, researchers gained more attention to the problem, and they proposed some robust method about face recognition. The main method is using the partial characters, but the use of whole characters is still little.This paper gives a systematically summarized introduction about subspace methods such as LDA, NMF and PCA; and robust estimations such as M estimate, L estimate and R estimate, and MCD which is an outliers detection method. Based on these, this paper uses an improved method which combined MCD with PCA to extract the whole characters of faces with outliers.The main purpose of MCD is to find a subset of vectors whose covariance matrix has the smallest determinant, and then use the iterative criterion and mahalanobis distance find the mean and the covariance matrix. After this, use the chi-square test to find out the outlier. Because of the problem of small sample, we use the good samples of the same class to replace the samples with outliers. Then give different weights to each sample, and calculate the final covariance matrix.In order to simulate the real abnormal disturbance of noise, this paper adds gaussian, salt and pepper and poisson noise to the selected samples. Test the method proposed in this paper on the ORL face database. The result of the experiment verified that the method has a good feasibility and validity.