基于改進(jìn)SwinTransformer的人臉活體檢測

打開文本圖片集

中圖分類號：TB9;TP391.4 文獻(xiàn)標(biāo)志碼：A 文章編號：1674-5124（2025）06-0031-09

Face anti-spoofing based on improved Swin Transformer

WANG Xuguang12，BU Chenyu12， SHI Zeyu1.2

（1.Departmentof Automation，North China Electric Power University，Baoding 07103，China; 2.Hebei Technology Innovation Centerof Simulation & Optimized Control for Power Generation，North China Electric Power University， Baoding ， China）

Abstract： With the development of facial recognition technology， face anti-spoofing as the security guarantee of facial recognition system becomes more and more important. However， the majority of existing face antispoofing models tend to concentrate on specific detection scenarios and atack methods，and have limited robustness and generalization capabilities when they face unknown atacks.For this reason， this paper proposes an improved Swin Transformer model， called CDCSwin-T（Central diference convolution Swin Transformer）. This model chooses Swin Transformer as the backbone to extract global facial information by using its shifted window atention.And it introduces the central difference convolution （CDC） module to extract local facial information. These enhance the model's ability to capture the diference between real and fake faces in this paper，thereby strengthening its robustness against unknown attacks; Additionally，the bottleneck attention module is introduced in the backbone model to guide the model to focus on the key information of faces and accelerate the training process.Finaly，the multi-scale information from different stages of the backbone model is fused byan adaptive fusion module，which improves the generalizability of the model proposed in this paper. The CDCSwin-T model achieves average classification error rates （ACER） of 0.2% ， 1.1%. （1.1±0.6）% ， and （2.8±1.4）% on the four protocols of the OULU-NPU dataset and the Half Total Error Rate （HTER） of 14.1% and 22.9% in cross-database evaluations on the CASIA-MFSD and REPLAY-ATTACK dataset， which both surpass the current mainstream models. It shows that its robustness and generalization capability when it face unknown attacks have been improved.

Keywords： face anti-spoofing; Swin Transformer; bottleneck atention module; feature fusion

0 引言

近年來，人臉識別系統(tǒng)性能不斷提升，其應(yīng)用范圍也逐漸擴(kuò)大，例如人臉門禁、人臉支付、考勤打卡、登機(jī)安檢等場景，但其帶來便利的同時(shí)也造成了一些問題。（剩余11616字）