Face Recognition Based On ICM Oscillation Time Series And Support Vector Machine

Face recognition is an important aspect of biometric identification, which has been widely used in various fields. But it is difficult to effectively extract the biological feature because of the influence comes from environment, illumination, angle and other factors which lead the characteristic to be unconspicuous. To achieve the feature extraction effectively and feature classification accurately has become a hot research topic at present.ICM model is very suitable to be applied in image processing which has the unique advantages such as simple structure, quick running speed, and it doesn’t need advance learning. Moreover, the oscillation time sequence of ICM used in image feature extraction has the characteristics like translational non-deformation, scaling non-deformation and rotation non-deformation. On the other hand, SVM is a kind of classifier with strong generalization ability, which can solve the problem of nonlinear, high dimension and so on. Therefore, this paper combined the advantages of the two methods, and put forward a algorithm combining the feature extraction based on oscillation time sequence of ICM with support vector machine.Firstly, this paper briefly describes the working principle and basic characteristics of the ICM model, deduces the dynamic properties of the model in detail, and discussed the characteristics of the period of the self-excitation and the acquisition period. Subsequently, according to the characteristics of ICM neuron which is the pulse synchronous release, the paper analyzes the geometric characteristics of the oscillation time series, and applies it in research of face recognition, and then proposes the feature extraction method based on ICM oscillation time series. Finally, the kernel function of SVM and the selection of the classifier are analyzed in detail, and the optimal one-to-one linear kernel function classifier is selected to classify the face feature.In the paper, a comparison experiment is conducted in ORL face feature database, which the optimal combinations obtained from experiments of conventional subspace methods respectively with different classifier are compared with the result obtained from experiment used method proposed. The simulation results show that compared with other subspace methods, the method proposed has superiority with high recognition rate, and the face image is less effect by translation, scaling, rotation. In addition, the method proposed is validated in the MIT-CBCL face feature database, and the results show that the algorithm has good feasibility in large data test space.