Construction and Implementation of Embedded-Based Face Recognition Models

This paper focuses on the efficient implementation of face recognition system under embedded platform, and proposes a set of systematic methods integrating feature extraction optimization, deep model compression and platform-level acceleration. A lightweight feature extraction structure with multi-scale convolution and attention fusion is designed to improve the expression capability of key regions; the model complexity is effectively reduced by techniques such as structural pruning, quantization-aware training and knowledge distillation.