1. BDC-DAL: A new algorithm using blueprint difference convolution and dual auxiliary supervision for lightweight face anti-spoofing(BDC-DAL:基于蓝图差分卷积和双重辅助监督的轻量化人脸活体检测算法)
- Author
-
叶继华(YE Jihua), 梁芳昕(LIANG Fangxin), 王超(WANG Chao), 肖顺杰(XIAO Shunjie), 宗义(ZONG Yi), and 江爱文(JIANG Aiwen)
- Subjects
face anti-spoofing(人脸活体检测) ,lightweight(轻量级) ,auxiliary information supervision(辅助信息监督) ,Electronic computers. Computer science ,QA75.5-76.95 ,Physics ,QC1-999 - Abstract
Face anti-spoofing is an integral part of securing face recognition systems. Most current face anti-spoofing methods employ deep convolutional neural networks, and although these systems can provide excellent detection performance, they often involve large number of parameters and computational complexity, which limits their application on resource-limited devices. To overcome these challenges, this paper proposes a lightweight face anti-spoofing method based on blueprint difference convolution and dual auxiliary supervision (BDC-DAL). This novel approach effectively reduces the parameters and computation of the model by combining a lightweight convolutional structure and blueprint difference convolutional design, while retaining the ability to capture complex fine-grained features. In addition, the introduced dual auxiliary supervision mechanism utilizes the reconstruction process of supervised signals to learn subtle spoof clues, which significantly improves the robustness of the model to unknown attack scenarios. To validate the effectiveness of the proposed method, this study conducts extensive tests on OULU-NPU and CelebA-Spoof, two recognized face anti-spoofing datasets. The test results are compared with those by performing the current mainstream deep learning face anti-spoofing algorithms and other lightweight network models. The experimental results show that the BDC-DAL method proposed in this paper dramatically reduces the model complexity, with the number of parameters reduced by nearly 90%; and the floating-point operations of the model reduced by about 10%. This shows that the BDC-DAL method successfully balances performance and resource consumption while achieving high efficiency and lightweight, providing a feasible solution for face recognition security in resource-constrained environments.(人脸活体检测是确保人脸识别系统安全不可或缺的一环。当前大多数人脸活体检测算法采用的是深度卷积神经网络,尽管其检测性能卓越,但往往涉及庞大的参数量和计算复杂度,限制了其应用。为克服这些挑战,提出了一种基于蓝图差分卷积和双重辅助监督的轻量级人脸活体检测算法(BDC-DAL)。通过将轻量化的卷积结构和蓝图差分卷积设计相结合,有效降低了模型的参数和计算量,同时保留了捕捉复杂细粒度特征的能力。此外,引入的双重辅助监督机制,利用监督信号重建过程学习细微的欺诈线索,显著提高了模型对未知攻击场景的鲁棒性。为验证算法的有效性,在CelebA-Spoof和OULU-NPU人脸活体检测数据集上进行了广泛测试,并与当前主流的深度学习人脸活体检测算法和其他轻量化网络模型进行了比较,实验结果显示,BDC-DAL算法大幅降低了模型复杂度,参数量减少了近90%,模型的浮点计算量也下降了约10%。BDC-DAL算法在实现高效性和轻量化的同时,成功平衡了性能与资源消耗,为在资源受限环境下的人脸识别提供了可行的解决方案。)
- Published
- 2025
- Full Text
- View/download PDF