Research On Locality Preserving Methods Based On The Measure Of Angel And Distance

The most fundamental problem for face recognition is feature extraction, at present most ofthese methods directly project the sample data to low-dimensional space, without using localstructure of the data information. People find that face image often presents a manifold structure,similar samples usually show an aggregate state, heterogeneous samples often show a scatteredstate. It was proved that manifold structure is essential for face recognition, so more and morestructure preserving algorithms have been proposed.In practice many structure preserving algorithms based on the distances between the sampleswhich describe the relationship of samples, then construct a close neighbor structure. This paperfound that when confronted with diversity data, the distances between the samples are often notgood to construct a close neighbor relations, but the angles between the samples as a measure tobe more effective. Thus, we propose the angle-based local preserving projections （ALPP）algorithm and the angle-based local discriminate projections （ALDP） algorithm based on previouswork.In order to construct more effective close neighbor structure, we introduce the complexfeature fusion mechanism. We combine the distance and angle through the complex vector form,and construct a complex close neighbor structure. After that, we propose complex local preservingprojections （CLPP） algorithm and complex local discriminate projections （CLDP） algorithm toensure that the projection of the close neighbor structure in the low-dimensional space not onlymaintain the original sample distance relationship, while original sample angel relationship,therefore the structure of the manifold could be better preserved.Experimental results on the public AR, FERET and COIL₂0databases demonstrate that ourproposed approaches are more effective than related methods in classification, they also provethat our approaches are more effective than related methods in retaining the local structure.