Feature Generation For Face Recognition

Face recognition as a special biometrics carries out identity recognition based on human facial feature information. Compared with fingerprint, iris and some other biometrics, face recognition is non-intrusive and can achieve satisfying performance without disturbing normal activities of people. Therefore, face recognition has a good prospect in the field of information security and public safety. However, in some circumstances, its accuracy is still difficult to achieve the ideal requirement, especially under the effects caused by changes of illumination, pose, expression and noise in the process of image acquisition. Mapping the original face image to a feature space by proper transformation to reduce the influence of these factors is an effective way to improve face recognition rate.This paper focuses on feature generation, using appropriate feature representation method to overcome the negative effect of illumination, expression and noise in face recognition. The highlights of this paper are as follows:1. Weber-based Local Multiple Patterns (WLMP) for face recognition:Encoding the amplitude of difference that is ignored in LBP code by using Weber Law. This method is simple to implement and of fast speed, and robust to pose variations, illumination and noise, but of poor robustness of expression.2. Multi-scale Block Local Multiple Patterns (MB-LMP) for face recognition:MB-LMP introduces some improvements to make up for the defects that exist in WLMP. This method combines the multi block method in MB-LBP with the coding method in WLMP. The algorithm shows robustness of pose variations, illumination and facial expression change. Meanwhile, it has a strong inhibitory effect on the influence of noise. When selecting an appropriate scale, this method even can completely suppress the influence caused by noise.3. Face recognition algorithm based on sparse representation:This part introduces the locality-constraint linear coding to sparse coding. The locality means that similar vectors can share similar code-words, thus preserving the correlation of coding. Applying the locality-constraint linear coding to joint features and the experimental results show that the method has a good effect on face recognition with interference of illuminations.