Research Of Lip Reading Algorithm For High Security Face Recognition Systems

With the wide use of face recognition systems,criminals begin to attack face recognition systems using forged or illegally obtained photos and videos.So it is urgent to design efficient liveness detection methods for high-level security.Aiming at defending against malicious attacks on face recognition systems,this paper proposes a liveness detection method by using lip reading algorithms.In order to overcome the problem of low recognition rate of traditional lip reading methods,this paper proposes three novel lip reading methods including a)HOG feature based lip reading method,b)stacked convolutional independent subspace analysis(ISA)based lip reading method and c)an end-to-end approach using deep neural networks.The main innovations of this paper include:Due to the lack of Chinese lip reading data,we have established a large Chinese digital lip reading database named YCLIP,which creates good conditions for the subsequent study of Chinese lip reading.After analyzed the limitations of manual designed features,we propose to use stacked convolutional independent subspace analysis algorithm,which is designed for video analysis,to extract lip features.Experimental results on YCLIP data set show that the stacked convolutional ISA algorithm performance better than the traditional HOG features.This paper presented an end-to-end approach for lip reading based on convolution neural network(CNN)and Long-Short Term Memory(LSTM)method.We use a AlexNet CNN to extract lip movement features and LSTMs to encode features from each frame.Due to the powerful ability of CNN and LSTMs,our method achieves amazing results on YCLIP and outperforms many published methods on public dataset MIRACL-VC.We obtain recognition rate of 70.0% in words recognition and 82.1% in phrases recognition,which are approximately 7% and 3% better than the best published results,respectively.The proposed lip reading algorithm performs the task of liveness detection well,making face recognition systems much safer.