Fusion of Edge‑Computing Technology for Smart‑Security Face‑Recognition System Development

Fusion of Edge‑Computing Technology for Smart‑Security Face‑Recognition System Development

Abstract—Traditional centralized face recognition systems suffer five critical engineering drawbacks: high network transmission latency, high risk of facial biometric privacy leakage, excessive cloud hardware and operation‑maintenance costs, service failure under network disconnection, and poor recognition robustness under complex illumination and occlusion conditions. In this paper, MobileFaceNet is taken as the backbone network. An ECA efficient one‑dimensional channel attention module is embedded to construct a CNN‑ViT hybrid feature extraction network. Structured channel pruning combined with 8‑bit layered mixed quantization is adopted to realize model lightweight deployment. A three‑level distributed collaborative system named “edge terminal‑edge node‑cloud management platform” is built. Core inference tasks including facial feature extraction and identity matching are offloaded to local edge‑device execution. Multi‑dimensional tests are carried out based on public LFW face dataset and self‑built security‑oriented occlusion dataset containing masked faces, sunglasses occlusions, large‑angle side‑faces and backlight samples. Complete ablation experiments are designed to verify independent gain of each optimized module. Horizontal comparative evaluation is performed against multiple mainstream lightweight face models. Measured experimental results show that the lightweight model achieves 98.7 % recognition accuracy under non‑occlusion conditions and maintains over 85 % accuracy under mixed‑occlusion scenarios. The model parameter volume is reduced by 32 %, floating‑point computation cost drops by 35 %, and storage footprint is compressed by 75 %. Single‑frame inference latency is controlled within 50 ms, yielding a 2.8‑fold speedup on edge devices. A single edge node supports concurrent processing of 48 video streams. Full recognition and local data storage can operate offline without network connection. Compared with conventional cloud‑centric schemes, cloud‑side hardware investment is cut by 70 %, bandwidth consumption decreases by 80 %, and long‑term operational cost falls by 65 %. The proposed algorithm and system jointly balance recognition accuracy, real‑time inference throughput and embedded‑hardware resource overhead. It can be deployed in residential compounds, industrial parks, transportation hubs and other security scenarios, and provides a reproducible, engineering‑ready optimization paradigm for edge‑oriented lightweight face‑recognition systems.

Keywords-Edge Computing; Smart Security; Face Recognition; Mobilefacenet; ECA Attention; CNN‑Vit; Model Lightweighting; Ablation Experiment