Article Preview
TopIntroduction
Identity verification is the determination of whether the operator has access to or uses certain resources. This, in turn, ensures reliable access. Identity verification plays a significant role in practical applications like speech interaction (Okumura et al., 2019), human-robot interaction (Freire-Obregón et al., 2021), secure driver identity verification (Ma et al., 2022), and attendance management (Akbar et al., 2018). Traditional identity verification consists of manual identification, IC card verification, password, and other methods. Although these methods are straightforward, they are characterized by low efficiency and easy imitation. Therefore, the methods cannot satisfy the current fast-paced social production and life.
Fortunately, as a result of advances in science and technology, biometrics like fingerprints (Grosz et al., 2022), iris (Yang et al., 2021), and face (Meng et al., 2021; Wang & Deng, 2021) etc.) have been applied to electronic identification. In practical scenarios, the three biometrics have their characteristics and requirements. More specifically, fingerprint-based identification technologies are susceptible to the cleanliness of the fingers. Some groups have few or even no fingerprint features, making it difficult to image. Iris-based identification technologies suffer from cumbersome detection processes and high hardware costs. Based on these concerns, this article investigates identity verification using human face biometrics.
Traditional identity verification approaches based on face biometrics rely on manual construction features (e.g., some geometric characteristics are constructed manually based on the eyes, nose, mouth, and other features of the human face.), which are then used to calculate certain statistical values or train a classifier to determine the final category. With the explosive growth of data volume and improvement of hardware level, deep learning (LeCun et al., 2015) has attracted attention for its excellent performance. It has also contributed to the development of the identity verification field.
Compared with traditional approaches, deep learning can automatically learn semantic-rich face features from data sets through a neural network without manual involvement. After years of experimental research, it has been confirmed that these features are more comprehensive and closer to the actual situation than the traditional methods. In addition, the generalization ability of the final model shows superior performance. More concretely, deep convolutional networks like deep ResNet (He et al., 2016), NFNets (Brock et al., 2021), and FisheyeHDK (Ahmad & Lécué, 2022) exhibit high accuracy on image classification problems. This is far more than the traditional machine learning approaches.
The most compelling characteristics of a convolutional network are local receptive fields and shared weights. These allow the network to automatically learn and form feature extractors and a classifier suitable for identity verification tasks. In addition, as the network deepens, its receiving domain expands, allowing convolutional kernels to capture more global information. This article leverages deep learning for identity verification on face biometrics. More precisely, the lightweight multitask cascaded convolutional networks (MTCNN) and MobileFaceNet, which are separately used for face detection and face recognition, are first analyzed (Chen et al., 2018; Guo et al., 2022; Zhang et al., 2016). Next, the overall architecture of Linux-based identity verification software is designed. Finally, the completed software is tested and demonstrated. The results show that the software function meets the design requirements and can complete the corresponding identity confirmation function.