Research On Face Anti-Spoofing And Recognition Technology Based On Neural Network

Facial detection and recognition technology plays an important role in security,access control,payment,public security,and other fields,because face features are important physiological information to mark the user’s identity.However,the limited memory and computational power of existing edge computing platforms cannot support real-time execution of relevant algorithms in traditional system architectures.Regarding the face anti-spoofing problem,some existing methods require additional multimodal devices,increasing system construction costs.On the other hand,the methods based only on RGB commercial cameras require the detected user’s behavior and actions,resulting in a lengthy verification process and poor user experience.Starting from the functional requirements of face anti-spoofing and recognition systems,this thesis investigates related research and selects appropriate methods to implement each function.To ensure that the system can run smoothly on a wider range of devices,this thesis prunes the face detection algorithm model to reduce model parameters and computational complexity while maintaining detection accuracy.For face recognition,a clustering strategy is introduced to effectively reduce the time for face searching in large-scale databases.For the face anti-spoofing task,this thesis proposes a silent face anti-spoofing model,which only requires an RGB face image as input and directly outputs the liveness verification result,providing good user experience and system response speed.Furthermore,this thesis constructs a face anti-spoofing and recognition system based on neural networks,which can be deployed and real-time run locally without network intervention,to some extent ensuring system security.The construction of the functional program is completed,and a client is developed to ensure simple and intuitive user interaction.The face detection algorithm is implemented and optimized,achieving an average detection time of 43.9ms,which is 81.2% of the original algorithm’s time.In face recognition tasks,after selecting an appropriate number of clustering clusters,the average matching time in a 100,000-face feature database is 17.1ms,representing 12.7% of the retrieval time without clustering when achieving an identification accuracy of 98.32%.In the face anti-spoofing task,the designed algorithm achieves an accuracy of 92.73% on a single dataset,with a decrease of 2.85%-5.49% in accuracy on cross-dataset validation,while traditional methods show a drop of over 14.76%.