Face Detection And Recognition Form Complex Background

Biometric recognition realizes personal identification based on the inherent physicalor behavioral characteristics of human being, among which facial features are the mostimportant and intuitional ones. Although the accuracy is lower than that of iris andfingerprint recognition, face recognition has become the most acceptable biometricrecognition method because of its no violation of privacy and intuition. Face recognitionis challenging in the pattern recognition and machine visualization due to the influence ofillumination|, expression and pose.In this paper, we present a comprehensive review of face detection and recognition,which includes the face detection, two-dimension and three-dimension face recognitionand the current development in the literature. We also give a presentation of the majorpoints and challenges in the face detection and recognition research. On the base of ourresearch, the discussion is organized as follows:(1) Discussing how to detect and recognize face from complex backgroundreal-time and efficiently only using gray scales.1. The famous AdaBoost algorithm is introduced and a proof is shown for theapproximate estimation in the training observation by using the AdaBoost algorithm. Anintuitional explanation for the AdaBoost algorithm is given that AdaBoost algorithm setsup a forward separate addition model to minimize the exponential loss function.2. Giving the proof for BP Neural network inverse transmit algorithm3. Introducing a commonly used integral image method.4. Designing a cascade classifier based on the AdaBoost and algorithm and themulti-layer local interconnected perceptron to detect the face and locate the facecomponents (2)In this paper, traditional elastic graph matching is analyzed and a face recognitionalgorithm based on local feature analysis and optimization matching is put forward.Firstly, some important face features (such as pupil, canthus, center of eyebrow, corners ofeyebrow, corners of mouth) are located using neutral network. Secondly, the multiscalefeatures of the feature points are extracted using the local mutiscale analysis feature of theGabor wavelet. In this way, every face feature point is represented by a series of Gaborwavelet coefficients. Finally, in order to find the face wanted, the test face is comparedwith the multiscale features of the corresponding feature points in the face database usingthe optimization matching. Here the optimization matching method is proved strictly. The test results on Yaleand ORL face database show the proposed method is not only far better than thetraditional EigenFace method but also overcomes the effect of the illumination variationon the face recognition and has quite good robust for face expression variation in somedegree. This method has three major advantages compared with elastic graph matching:firstly, it fully uses commonness of the feature point in the linear subspace spanned by itsfeature vectors, which can not be realized by elastic graph matching; secondly,tremendous computation has been reduced using the neutral network; thirdly, we can getthe global analytic optimum solution using the optimization matching at the feature points,which avoids lots of iterative computation (sometimes iterative computation can only getsuboptimum solution) and improves the matching precision and recognition accuracy aswell. (3) The performance of face recognition system is greatly affected by theillumination changes. In this article, we propose a method of face illuminationcompensation from rough to accurate and from whole to local based on neural networkand wavelet. From rough to accurate means from the rough linear illuminationcompensation in homomorphism region to the accurate illumination compensation inneural network and wavelet. From whole to local means the illumination compensationfrom the lower frequency (whole) to the medium and higher frequency (local). This method sufficiently combines multi-resolution analysis of wavelet and the self-adaptation learning and good spread ability of BP neural network, thus this method carriesout the face illumination compensation. Compared with other methods in the literature,this method has some distinguishing advantages. It doesn't need to estimate theparameters of the face surface (such as the reflection rate, normal vector etc). Neitherdoes it need to estimate the direction and intensity of environmental light nor the lighttypes (such as point light, parallel light, single light source and multi light source). At thesame time, non-face images can not be changed into human faces after the compensation.