Research On The Key Techniques Of Multi-mode Face Anti-spoofing

Face anti-spoofing is a research direction that has received much attention in the field of computer vision in recent years.A large number of face image identification technologies have been deployed in application scenarios such as digital office and financial technology.In these scenes,different environmental backgrounds,lighting conditions,and face poses will cause significant differences in face images.Faces images displayed on tablets and other devices are easily recognized as living faces,which poses serious security risks.Face anti-spoofing have an urgent need in these application scenarios.Although some progress has been made on face anti-spoofing,there are still some problems that need to be solved.First of all,the traditional face anti-spoofing algorithm is sensitive to face gestures,recognition rate is low,and the user experience is poor.How can we improve the recognition accuracy in face anti-spoofing problems,get rid of the steps of user cooperation,and improve the user experience;second,In the face antispoofing task,how to make full use of limited data and augment the data in the absence of relevant data sets;finally,how to effectively use the characteristics of multiple modal data under multi-modal data research.This article will focus on the above three problems and conduct research on the multi-modal face anti-spoofing problem.The main work of this article is as follows:(1)Research on face anti-spoofing algorithm based on RGB multi-modal generation.Since RGB images are currently the most common images and the most widely used modalities,this paper proposes a face anti-spoofing method that generates multi-modal features from a single RGB image,which is generated from the RGB images of the currently public data set Different modalities,such as high dynamic range(HDR)images,depth maps,etc.,and then study the impact of different modalities generated by the analysis on the face anti-spoofing task.(2)Construction of large-scale multimodal face anti-spoofing dataset.Aiming at the problem of lack of available multi-modal human face detection data set in the academic world.This paper constructs a large-scale multimodal biopsy data set.The dataset collected 50 real-person videos in 10 scenes and attack videos in 11 scenes.Segmented video is recorded simultaneously with wide dynamic range images and near infrared images.It's a total of 2100 videos,the original video frame number exceeds 5 million frames.(3)A multi-modal face anti-spoofing algorithm is proposed and a face anti-spoofing system is designed.According to the dataset proposed in this paper,a multimodal network model for face anti-spoofing is proposed.This method combines the information from different modes to do face anti-spoofing algorithm.Then according to the algorithm,a GPU accelerated detection system is implemented based on Python language with Pytorch,Opencv and other open-source algorithm libraries,which provides the necessary research and analysis basis for researchers.The algorithm framework proposed in this paper realizes the target of living detection.Experiments on the open dataset show that the method of generating high dynamic range images proposed in this paper is better than other algorithms in this paper.And the multi-modal living detection algorithm constructed in this paper has significant advantages over the single-mode method in the multi-modal face anti-spoofing dataset and has a better application prospect.