Abstract
As social awareness of personal and public safety grows, face recognition technology has become increasingly integrated into daily life. This paper presents a face recognition system based on a Deep Residual Network (ResNet), implemented across both PC and Android platforms. The PC terminal features a voice broadcast function to provide real-time alerts, while the Android implementation utilizes the Scripting Layer for Android (SL4A) to harmoniously integrate Python-based deep learning models with the mobile operating system. Experimental results demonstrate that the system achieves high recognition accuracy, providing a portable and effective solution for identity authentication and secure mobile operations.
Introduction
Face recognition has transitioned from specialized security applications to common consumer technology, appearing in airport checkpoints, mobile payments, and unmanned retail environments. However, deploying high-accuracy models across diverse hardware remains a significant challenge. Traditional Convolutional Neural Networks (CNNs) often face the "vanishing gradient" problem as they increase in depth, where the training signal weakens before reaching early layers, limiting their accuracy.
Furthermore, there is a persistent trade-off between computational power and portability. While PCs offer the necessary resources for complex models, they are not easily portable. Conversely, mobile devices offer portability but typically require different programming environments (e.g., Java or Kotlin), complicating the deployment of research-grade Python models. This research bridges this gap by creating a unified, high-accuracy system for both environments using a Deep Residual Network architecture.
Methodology
The core of the proposed system is a 34-layer Deep Residual Network (ResNet). Unlike standard CNNs, ResNet incorporates "shortcut connections" or skip connections that allow gradients to flow more easily through the network. This architecture enables the extraction of abstract and highly distinguishable facial features, outperforming older architectures like VGG.
Technical Implementation
The system's technical stack is designed for cross-platform compatibility:
- Dlib: Utilized via a Python interface for face detection, identifying facial key points, and extracting feature descriptors.
- OpenCV: Employed on the PC side for real-time video capture from USB cameras.
- SL4A (Scripting Layer for Android): This acts as the bridge for mobile implementation. It allows the Python-based recognition logic to run on Android by serving as a Remote Procedure Call (RPC) agent, enabling the Python script to access the Android camera and display results.
- Euclidean Distance: Identification is performed by calculating the mathematical distance between a captured face's features and a database of known candidates. A match is confirmed if the distance falls below a precision threshold of 0.4.
Results and Discussion
The ResNet-based model demonstrated significant discriminatory power and efficiency. In benchmark evaluations, the underlying residual network architecture achieved a low error rate of 3.57%. In real-time system testing, the recognition accuracy exceeded 96%.
Mobile Terminal Performance
The "grafting" approach using SL4A proved highly effective. By porting the PC-developed Python code directly to the mobile terminal, the researchers ensured consistency in recognition logic while drastically reducing development time. The Android implementation successfully performed real-time detection and recognition, displaying results directly on the mobile interface.
In a specific test case involving seven candidates, the system correctly identified "Liang Juan" by finding the minimum Euclidean distance among the gallery images, confirming the effectiveness of the chosen 0.4 threshold for filtering incorrect matches.
Conclusion
This research successfully demonstrates a portable, high-accuracy face recognition system. While the current implementation shows excellent results, the authors acknowledge that environmental factors such as lighting conditions, photo quality, and physical obstructions (like hats or glasses) can still impact performance. Future work will focus on enhancing the system's robustness against these variables to ensure reliable operation in uncontrolled, real-world environments.
Summary updated