Lightweight face recognition network design and performance optimization based on embedded AI platform

With the growth of artificial intelligence technology, lightweight face recognition methods have shown great application potential in many fields. However, when existing face recognition networks are deployed on embedded platforms, it is often difficult to balance accuracy and efficiency due to high model complexity. Existing lightweight methods still have problems, such as significant accuracy loss and poor cross-scenario adaptability. To enhance the accuracy, computational efficiency, robustness, and practicality of lightweight face recognition methods, this study proposes a lightweight face recognition network model based on improved MobileFaceNet. To address the aforementioned challenges, this study proposes a holistic lightweight face recognition network based on an enhanced MobileFaceNet architecture. Unlike mere architectural tweaks, a training-inference decoupled design is introduced through branch fusion, a multi-objective collaborative optimization loop incorporating triplet loss and symmetry-aware regularization, and a knowledge distillation mechanism for cross-model generalization. These components work synergistically to achieve an optimal balance between accuracy, computational efficiency, and deployment feasibility on embedded AI platforms. Experimental results on the LFW and WIDER FACE datasets confirms that the proposed method achieves a better accuracy-efficiency trade-off than several state-of-the-art lightweight benchmarks. This paper is based on the MobileFaceNet and optimizes the network structure by constructing a forward inference path and a multi-objective collaborative training optimization loop. It reduces the computational complexity and resource usage of the model by integrating lightweight methods such as triplet loss and a knowledge distillation mechanism, and finally deploys the optimized model on the embedded AI platform. Experiments showed that the overall size of the proposed method was less than 10MB, and its overall misrecognition rate was less than 20% under challenging test conditions (e.g., with low confidence thresholds or in scenarios with significant occlusion and illumination variations). In typical operational settings, the threshold could be tuned to achieve a far lower misrecognition rate while maintaining a reasonable rejection rate. Compared with existing lightweight approaches, the proposed method achieves a better trade-off between accuracy and model complexity, maintaining high recognition performance while significantly reducing computational overhead and memory footprint on embedded platforms. In practical application tests, the overall memory usage of the proposed method under different model configurations was less than 200MB, and the inference time was less than 200 ms when the input image size was less than 11 pixels. These results demonstrate that the proposed method outperforms state-of-the-art lightweight face recognition networks in terms of both efficiency and practicality, especially under resource-constrained environments. Individual techniques such as branch fusion, PReLU, triplet loss, symmetry loss, and knowledge distillation are well-established. However, their integration into a unified lightweight framework that enables training-inference decoupling, multi-objective collaborative optimization, and efficient cross-model generalization represents a novel systematic design paradigm for embedded face recognition.