Study On Face Recognition Based On Subspace Analysis And Frequency Domain Feature Extraction

The face recognition technology has been widely applied in the field of publicsecurity management, intelligent monitoring, digital identity and digital entertainment,which has brought great convenience to human production and life. However, most facerecognition systems today are working under certain restrictions, because facerecognition is vulnerable to the impact of factors such as illumination, facial expression,posture changes, and face recognition generally suffers from high-dimension dilemmaand the small sample size problem, so how to overcome these deficiencies isparticularly crucial. This paper mainly expands around these two issues and overcomingillumination changes, facial expressions, and other factors. Summarized as follows:1. Expound the face recognition technology, why to use face recognition and someface recognition applications. Summarize the domestic and external research progress offour kinds of face recognition technology, which are the algorithm based on frequencydomain, linear subspace analysis, nonlinear subspace analysis based on kernel mappingand manifold learning algorithm.2. The analysis and discussion are focused on the contacting between the lineardiscriminant analysis, and its extension methods, and reach the following conclusions.Any discrimination information will never be lost when projecting onto the non-zerospace of total scatter matrix. The most extension methods of linear discriminant analysisare either using only null space of within-class scatter matrix, or utilizing only non-zerospace of between-class scatter matrix, leading to take no full advantage of all thediscrimination information. Maximum margin criterion algorithm can effectivelyovercome the small sample size problem, and its operation is simple, and is a veryefficient and stable algorithm.3. The face recognition algorithm based contourlet transform and coupled subspaceanalysis was presented, firstly faces are decomposed by contourlet transform, pluralfrequency domain coefficients in the same scale and different directions are fused into asubband, then face features are extracted in fused subbands by improved coupledsubspace analysis, finally the nearest classifier is adopted. Experiments on commondatabases show that our method is effective and has a higher recognition rate comparedwith the principal component analysis and coupled subspace analysis.4. A method called uncorrelated locality preserving projection analysis based on effective and stable maximum margin criterion is proposed. The advantages anddisadvantages of locality preserving projections and maximum margin criterion arediscussed, then new classification criteria included between-class and within-classlocality information is gained. Dimensionality reduction is achieved through singularvalue decomposition, the analysis found that the discriminant feature set based on themaximum margin criterion is generally statistical correlated, which are not favorable forthe correct recognition, so the method based on uncorrelated discriminant featureextraction is derived. Experiments on several databases show that the method iseffective and stable and has a higher correct recognition rate compared with the localitypreserving projections, linear discriminant analysis and locality-preserved maximuminformation projection algorithm.