Research On Deep Learning-based Face Liveness Detection

Facial liveness detection technology is an important method used to verify human identity.It has multiple significances,such as enhancing security,preventing fraud and theft,protecting personal privacy,improving user experience,promoting industry innovation,and advancing technological progress.Despite the widespread application and development of deep learning-based facial liveness detection technology,the current methods still suffer from high parameter complexity and excessive redundant information.Additionally,existing mainstream models primarily utilize single-modal information,resulting in low detection accuracy and poor robustness.To address these issues,this paper proposes improvements to facial liveness detection methods based on deep learning.The specific contributions are as follows:1.To tackle the problem of high parameter complexity and excessive redundant information in conventional convolution,this paper introduces an asymmetric convolution-based facial liveness detection method called FASNet.By incorporating asymmetric convolution,the model’s receptive field is increased without adding parameters,reducing redundant information,and thereby improving the accuracy and speed of facial liveness detection.To further enhance the precision of the improved FASNet model,the CBAM attention mechanism is incorporated,effectively enhancing model performance.By combining the advantages of these two methods,the accuracy and efficiency of the FASNet model are significantly improved.Experimental results demonstrate that this method achieves favorable performance on multiple datasets,with an average minimal classification error rate of 1.27%on the mainstream facial liveness detection dataset Replay-Attack.2.To address the issues of poor robustness and insufficient accuracy in mainstream models,this paper incorporates the SENet attention mechanism into the Swin-Transformer model.Experimental results show that the improved algorithm effectively enhances the accuracy of facial liveness detection.Additionally,multiple modalities(IR,RGB,Depth images)are employed to capture facial information in the improved facial liveness detection model,enabling accurate recognition of various attack methods and enhancing the model’s robustness.Experimental results demonstrate that this method achieves good performance on different datasets and multiple models.On the facial liveness detection dataset CASIA-SURF,the average minimal classification error rate reaches 3.13%.3.Facial liveness detection system.Based on the research and improvements in the aforementioned aspects,an efficient and accurate facial liveness detection system is implemented using the PyQt5 framework.Firstly,key technologies of the facial liveness detection system are introduced.Secondly,a requirement analysis is conducted from both functional and non-functional perspectives.Next,the system’s various functions are described.Finally,the development of the facial liveness detection system is completed.