Lightweight deep learning research is an essential technology for embedding existing cloud-based learning models into lightweight devices, providing various benefits such as reduced latency, protection of sensitive personal information, and reduced network traffic[1]. Because centralized cloud-based GPU-based computing resources have high costs and power consumption, decentralization allows efficient use of computing resources while performing desired functions by developing and applying lightweight deep learning models in an embedded environment capable of distributed processing. In addition, in the field of facial recognition security, non-invasive biometric methods such as facial recognition reduce the risk and difficulty of processing confidential biometric data. It also simplifies the task of providing security while ensuring accurate results. Integrating biometric systems with edge computing and deep learning reduces latency and bandwidth usage, making the system more robust and dynamic[2].

In this study, we compare two lightweight deep learning models for face recognition suitable for the mobile environment and propose a more suitable model.

MobileFaceNet is a model optimized to deploy it in embedding environment and is used by many companies as a commercial face recognition model[3].

On the other hand, ResNet, which is being compared in this paper, is known as a network that shows high performance by sequentially stacking residual blocks through skip connection. In addition, ResNet-100 was used to simultaneously consider face recognition performance, speed, and memory through learning and evaluation.

In this paper, we compare the performance between the two models and propose a lightweight deep learning model suitable for the mobile environment with improved recognition accuracy for face recognition and security.

This paper is organized as follows. In Section 2, we describe the dataset and preprocessing to be used in the experiment. In Section 3, we examine the characteristics of the two models compared in this paper. Finally, in Section 4, we present the experimental results between the two models and suggest a model that is more suitable for the mobile environment.

In this chapter, we will explain the learning data set selection and preprocessing methods, which are important for the performance of face recognition models.

2.2 Data preprocessing

Various methods for face detection and alignment have been proposed for preprocessing of learning data.

As shown in Fig. 1., An overview of the single-shot multi-level face localisation approach(a). Detailed illustration of RetinaFace design(b). RetinaFace is designed based on the feature pyramids with five scales. For each scale of the feature maps, there is a deformable context module. Following the context modules, we calculate a joint loss (face classification, face box regression[5], five facial landmarks regression and 1k 3D vertices regression) for each positive anchor. To minimise the residual of localisation, we employ cascade regression[3]. five landmarks (eye centers, nose tip, mouth corners) were extracted through RetinaFace[5], and the faces were aligned using the affine conversion function provided by opencv[6].

Fig. 1. 
RetinaFace structure used for five landmarks[5]

This paper used the WebFace42M dataset, which is the largest training set for face recognition. For face detection and alignment, five landmarks were extracted through RetinaFace and the faces were aligned using the affine conversion function provided by openCV.

In experiment result, as shown in Table 1, the performance of the ResNet-100(SE-LResNet100E-IR) model is slightly better. The biggest difference was shown in 'Recall rate', which is the part that accurately recognizes the existing face value in facial recognition security and is the most important part in biometric information analysis. 'F-1 score' and 'Accuracy' showed the next difference, and 'Precision' and 'Specificity' showed a slight difference in the performance for misrecognition.

There has been a lot of development in the technology of face recognition. In particular, improving the accuracy of Edge AI is an important part of facial recognition security, and it can simplify the task of providing security through accurate results[2]. In this paper, the performance comparison results between the two models show that the ResNet-100 (SE-LResNet100E-IR) model shows slightly better performance in Accuracy, Precision, Recallrate, Specificity, and F-1 score. This is expected to be more meaningful in limited embedded environments.

We did not consider computational efficiency through model optimization in the lightweight deep learning model.

There is no data that directly compares the hardware usage environment, such as computational resources and memory usage, between the two models presented in this paper, and this requires additional research. It is also necessary to study the optimization of the model structure by considering the number of parameters of the lightweight deep learning model, FLOPs (Floating point Operations Per second), etc.[9], so in the future, we plan to study a model that balances computational efficiency and accuracy well through model structure optimization by considering various embedded environments.