GhostFaceNet++: boosting efficiency and accuracy via CSP bottlenecks and Channel Attention

Abstract Deploying accurate face recognition systems on resource-constrained devices remains challenging due to the high computational demands of many state-of-the-art models. This paper proposes GhostFaceNet++, an improved lightweight framework that enhances the GhostFaceNet architecture to achieve both efficiency and accuracy. The framework incorporates a Cross Stage Partial (CSP) structure into the GhostNet bottleneck to enrich feature representation and reduce computational redundancy, while an Efficient Channel Attention (ECA) mechanism is integrated into the classification head to enhance discriminative feature learning. Extensive experiments on standard benchmarks (LFW, CFP-FP, AgeDB-30, CP-LFW) demonstrate that GhostFaceNet++ consistently outperforms the original GhostFaceNet and achieves competitive performance with state-of-the-art lightweight models. For example, the ECA-CSP GhostFaceNet V1-2 variant improves AgeDB-30 accuracy from 89.80 to 90.67% when trained on CASIA-WebFace, while reducing FLOPs from 60.3 to 51.6M and model size from 8.13 to 7.51 MB. GhostFaceNet++ operates in the highly efficient 51–62 MFLOPs range, outperforming the original GhostFaceNet (60–275 MFLOPs) and being significantly more efficient than MobileFaceNet (439.7 MFLOPs). These results confirm that the integration of CSP and ECA enables GhostFaceNet++ to strike a favorable balance between compactness and performance, advancing the design of lightweight face recognition systems for real-world deployment on resource-constrained devices.