Study On Illumination Preprocessing And Feature Extraction For Face Recognition

Face recognition,as a user-friendly and non-contact identity verification and recognition technology,has great application potential in human-computer interaction,information security and entertainment.It has been a research spot of computer vision and artificial intelligence,and has extracted the attentions of many researchers since the 1960 s.After decades of study,significant progress in this field has been achieved,and some face recognition systems are already commercially available.However,designing an effective face recognition system still faces several urgent problems that must be tackled,including illumination variation,running on resource-restricted devices(e.g.,smart phones),and spoofing attacks.This dissertation mainly concerns three aspects of face recognition,which are illumination preprocessing,face representation,and face spoofing detection.The main contributions of this dissertation are the following:(1)A wavelet-based image preprocessing method for illumination insensitive face recognition is proposed.After applying single-level 2D discrete wavelet transform,how illumination variations affect the four sub-bands is studied.The observation is that most of the variations exist in the low frequency sub-band while the three high frequency sub-bands are almost unaffected.Therefore,this dissertation proposes to process the low and high frequency sub-bands separately.For the former part,a chain of histogram equalization,locally adaptive gamma correction,difference of Gaussian filtering,and contrast equalization is applied.For the high frequency part,unsharp filtering is used to enhance edge information.On three publicly available databases,the proposed method provides better performance for illumination normalization than existing methods in the literature.(2)Directional gradients in the logarithmic domain are integrated to form the illumination invariant.Based on the Lambertian reflectance model,a face image is the product of face reflectance and incident light.Face reflectance is the intrinsic property of a face,and thus it is independent from illumination.Therefore,the reflectance component is very suitable for illumination insensitive face recognition.The logarithm transform is firstly applied to a face image such that the multiplicative model is converted to an additive one,which could make the extraction of illumination invariant relatively easier.Based the common assumption that the incident light component remains the same for different face regions while the reflectance component varies abruptly,the gradient of a logarithm image merely contains the reflectance component.Finally,the anisotropic diffusion is utilized to integrate the directional gradients,generating the final illumination-invariant face image.Experiments on three publicly available databases show that our method can significantly improve the robustness of a face recognition system against illumination variations.(3)A gradient-based texture descriptor is proposed for face representation.Texture feature has been widely used in the field of face recognition due to its good descriptive power and low computation complexity.Because it has been observed in some research works that gradient-based texture features are more discriminative than texture features in the spatial domain,texture features are extracted from the gradient maps of a face image.Moreover,weights are assigned to different local face regions,according to the discriminative capacity of the information in local regions.These weights are used to construct a weighted similarity index between two face descriptors.On three public face databases,our method shows better performance than previous state-of-the-art texture-based face descriptors.(4)A local dynamic texture descriptor is designed for two dimensional face spoofing detection.There is little attention that has been paid to face spoofing attacks until recently.Among multiple types of spoofing attacks,two dimensional attacks,in which an imposter presents a photo or play a video in front of a face recognition system to illegally obtain the access right of some valid user,are usually low-cost and easy to conduct.Thus,it is very urgent to tackle this type of attack.Considering that the attack media is quite different from a real face,including differences in micro-motion and texture characteristic,the local binary count patterns in a video clip are analyzed to capture these differences,and further distinguish a real face from a falsified one.Experiments on three public face anti-spoofing databases show that our method achieves very promising performance of countering spoofing attack,outperforming many state-of-the-art methods in the literature.