Design Of Face Anti-spoofing Based On Deep Learning

Recently,deep learning has dramatically improved the state-of-the-art performance of many computer vision tasks and brought huge development prospects.Face recognition technology has also made a breakthrough,and face recognition systems become the most widespread biometric applications on the market.However,face recognition systems have potential risks.Criminals will use various methods to carry out fake attacks,impersonating the identity of the target to complete illegal acts.To ensure the safety of the face recognition system,face anti-spoofing is a crucial internal component of the system.Aiming at face anti-spoofing,this thesis conducts explorations from algorithm research and engineering application,including face anti-spoofing algorithm with RGB cameras,adaptive multi-modal fusion for lightweight face anti-spoofing,face antispoofing application systems.The main work and contributions of this thesis are as follows:1.In this thesis,the learnable wavelet decomposition module is designed to effectively separate the features of different frequency bands and enhance the spoofing cues on different frequency bands.Besides,as diverse types of fake faces and large inter-class differences in the embedding space,we also design a single constrained center loss to compresses intra-class variations of live faces,and the network can learn more discriminative features with less optimization difficulty.2.In this thesis,an adaptive multi-modal fusion method is proposed for the task of multi-modal face anti-spoofing.The weight is redistributed through the attention mechanism,and suitable modal representations are selected for different characteristics of face presentation attacks.And it is the first attempt to introduce the idea of central differential convolution into the deep separable convolution,so that the lightweight model can learn more detailed discriminative features.3.Finally,this thesis combines the research and engineering experience in face anti-spoof,respectively design and implement the face anti-spoofing system for RGB cameras and the face anti-spoofing system for Depth cameras.These face anti-spoofing systems have achieved excellent results in their respective test sets.