The Research Of Face Recognition Algorithm Based On Sparse Representation

Face Recognition (FR) is considered as an important technology of modernbiological information recognition. Compared with other Biometric identificationtechnology, it has more advantages such as intuition, simplicity, extendibility and so on.As for this, FR is widely developed and applied in recent years. What we say about FRis that for a given face image, confirm the identity of the person by using the storedface database. As given in the reference, most of the existing FR methods needcomplex preprocessing and feature extraction. In complex condition, the choice offeatures has weak robustness. What’s more, different choices may have great effect onthe recognition rate. All these problems make the existing FR methods restricted in thepractical application.Sparse representation is a novel algorithm, which has high recognition rate andstrong robustness and attaches quiet a lot of researchers. In this paper, the fundamentaltheory of the sparse representation is introduced at first, and then more energy is givento the usage in the field of FR. In chapter III, the classical sparse representation methodis depicted. Moreover, we analyze the feasibility and advantage through theexperiments. Based all of this, a novel sparse representation algorithm, the basicthought of which is choosing the proper training samples to reconstitute the testingsample, and classifying the testing sample by the result of the reconstitution, isproposed. In chapter IV, we select only one training sample from each class, whichcompose a new training sample set. From the experiment results we can easily find thatthe similarity of the two samples affects the recognition directly. In other words, themore similar of the two kinds of sample has, the higher recognition rate the result has.As for this, in order to obtain the more accurate result, we choose the nearestneighbors for testing sample from all the training samples. In order to deal with thecomplex condition such as illumination, shelter etc. We introduce the Gauss kernelFunction in the subspace of the face image for the choosing process. Compared with chapter IV, the training samples we get have much more similarity with the testingsample. A number of experiments show that our algorithm can achieve betterclassification and less time complexity.