Problem
Research questions and friction points this paper is trying to address.
Mobile Platform
Facial Recognition
Compact Model
SqueezeFacePoseNet: Lightweight Face Verification Across Different Poses for Mobile Platforms
🤖 AI Summary
To address the challenges of low accuracy and model bloat in mobile cross-pose face recognition, this paper proposes a lightweight pose-robust recognition method. Our approach introduces three key contributions: (1) an ultra-lightweight CNN architecture derived from SqueezeNet; (2) a pose-aware channel pruning strategy coupled with multi-view feature alignment, enabling pose-invariant feature extraction under stringent parameter constraints; and (3) an integrated optimization framework combining pose-normalization loss, knowledge distillation, and INT8 quantization for efficient deployment. Evaluated on the CFP-FP and IJB-C cross-pose benchmarks, our method achieves true positive rates of 98.2% and 92.7% at FAR=1e−3, respectively. The resulting model occupies only 1.2 MB and attains 42 FPS inference speed on an ARM Cortex-A72 processor—demonstrating an exceptional balance among accuracy, computational efficiency, and on-device deployability.
Innovation
Methods, ideas, or system contributions that make the work stand out.
SqueezeFacePoseNet
LightweightFaceRecognition
MobileOptimizedModel
🔎 Similar Papers
No similar papers found.
