Face Spoofing Detection Based On Image Feature Analysis

With the continuous development of face recognition technology in recent years,face recognition has been widely used in daily life.Accordingly,the risk of face spoofing attacks in face recognition systems has been greatly increased.To this end,this thesis addresses the issues in face spoofing detection and tries to improve the security of face recognition systems.The main works of the thesis are as follows.By analyzing the differences between genuine faces and spoofing faces in sampling process and reproduction process,a face spoofing detection algorithm,which integrates multiple image features in spatial and frequency domains,is proposed.Firstly,the chromatic moment features and blurriness features are extracted from face images in spatial domain,and the energy features and the energy percentage features of image Fourier spectrum are extracted in frequency domain.Then,these features resulted from spatial and frequency domains are cascaded and normalized as a holistic feature vector to represent a face.Finally,the feature vectors are used to train a SVM classifier and to distinguish genuine faces from spoofing faces.Experimental results on two pubic datasets,i.e.,NUAA and CASIA,indicate that the proposed algorithm is superior to other algorithms based on single feature or single domain features in face spoofing detection.To overcome the problem that the existing face spoofing detection algorithms need to be elaborately designed,and the face spoofing detection algorithms based on conventional deep learning lack sufficient training samples,a face spoofing detection algorithm based on deep transfer learning is proposed in this thesis.Firstly,the VGG-16 convolutional neural network model is modified to make the network structure more suitable for face spoofing detection.Then,data enhancement and transfer learning are used in deep network training.The images of training set are preprocessed and sent to the network for fine-tuning model parameters.Finally,the trained model is used to predict a face image in order to achieve face spoofing detection.Experiments are also conducted on the two public datasets NUAA and CASIA datasets,and the results indicate that the proposed algorithm can further improve the performance of face spoofing detection.