Research Of Embedded Face Recognition Algorithms Based On Curvelet Transform

Since "9.11", the terrorists’ attack, there is an urgent demand of a fast andeffective authentication technique to ensure system security and public safety. Theauthentication method based on face recognition has brought hope to people. Facerecognition is more direct, friendly and convenient than other biometric identityrecognition method, is more acceptable for people. Face recognition in the earlyperiod only runs on a PC with good performance, due to high dimension and largescale of data. With the development of the face recognition technology and embeddeddevices (especially ARM devices), face recognition can run on the embedded devices.Embedded system is easy to carry, quickly to install and of strong mobility, hasreplaced the PC in many fields. The research and implement of face recognitionalgorithm in this article is actually based on the embedded platform of PXA270.Wavelet transform has a good capability of temporal and spatial localization. Itcan represent the point singularity of signal better. However, we often focus on thecurve characteristics of the edge of the facial contours and five sense organs whenface recognition. The supporting basis of the Curvelet transform is anisotropy, whichmakes the Curvelet transform reflect the characteristics of along the edge directionbetter. Thus this paper uses the Curvelet transform for facial feature extraction. AsWavelet transform has disadvantages in face recognition and the resources ofembedded devices are limited, this paper studys two embedded face recognitionalgorithms: embedded face recognition based on Curvelet and2DPCA, embeddedface recognition based on Curvelet feature weighted fusion. In the first algorithm, wefirstly use Curvelet transform to extract low frequency coefficients containing themost energy of the face, then use2DPCA to reduce the feature dimension andcomputational complexity, and finally recognize and classify face combined withnearest neighbor classifier. In order to highlight the contribution of the Curveletlow-frequency parts to face recognition and not abandon the role of high-frequencypart, this paper proposes a embedded face recognition algorithm based on Curveletfeature weighted fusion. First use Curvelet transform to extract facial features atdifferent scales and in different directions, and use2DPCA to reduce the dimension ofthese features, then calculate the weight values of these local features at differentscales and directions referring to the idea of the Fisherface algorithm, finally calculate Euclidean distances between these local features and face training samples, weightand fuse these distances together as a final distance value to recognize and classifyhuman faces.Because resources of ARM platform is limited and the dimension of the faceimage is very large, the training process of face classifier is done on the PC and sentto the ARM platform with the network to recognize face. After the test based on ORLand Yale face databases, we finds that embedded face recognition algorithm presentedin this paper can run well on embedded platform and it gets a good recognition effect,shortens the recognition time. Meanwhile, the proposed algorithm can effectivelyreduce the impact of the lighting conditions and changes in the expression on the facerecognition. The research in this paper has some theoretical value and reference forface recognition and development of embedded systems.