Research on Community Policing Management Based on ECA Lightweight Face Recognition Models

Yan Wang
Jilin Police College ,130000, Jilin, China

Abstract:

This paper proposes a lightweight ECA attention mechanism to enhance facial recognition performance. The face is segmented into 7 regions, weighted differently, and optimized using one-dimensional convolutional channel attention. Adaptive mean pooling integrates spatial information, optimizing facial feature vectors via cosine loss. This technology enables efficient community personnel management, supported by automation and information technology. In ablation experiments, the ECA module achieved high recognition rates of 99.49% on LFW and 95.27% on Agedb_30, with 12.8ms speed. It also achieved 97.9% accuracy on a 500-person dataset. Comparing three attention models, convergence was achieved after 8000 iterations, stabilizing accuracy at 99%, validating the effectiveness of the proposed management system.