FD-GhostFaceNet: Frequency-Decoupled Ghost Modules for Lightweight Face Recognition

Deploying accurate face recognition models on resource-constrained devices remains a significant challenge. Ghost modules mitigate feature-map redundancy by synthesizing half of each feature map through inexpensive linear transformations and form the basis of the lightweight GhostFaceNet and GhostFaceNet++ families. Nevertheless, the inexpensive generator is a strictly local 3 × 3 depthwise convolution: the synthesized features remain near-duplicates of the intrinsic ones, and deeper stages lack the global receptive field required for cross-pose and cross-age matching. This paper introduces FD-GhostFaceNet, built upon the proposed Frequency-Decoupled Ghost (FD-Ghost) module, which replaces the local generator with a complementary-band global counterpart. A learnable, input-conditioned partition of the two-dimensional Discrete Cosine Transform spectrum decomposes the intrinsic features into a global low-frequency band that encodes pose- and age-stable identity structure and a locally refined high-frequency detail band. Nyquist-consistent anti-aliased downsampling, which band-limits each feature map before subsampling so that the decoupled high-frequency band survives decimation, completes the frequency-decoupled design. Applied to both published trunks, FD-GhostFaceNet-V1-2 and FD-GhostFaceNet-V2-2 require 4.60 M and 7.39 M parameters, 83.1 and 97.1 MFLOPs, and 9.19 and 14.77 MB of storage, respectively, and thus fall within the sub-100 MFLOPs category of lightweight face recognition models. Across six standard benchmarks, both variants surpass their corresponding GhostFaceNet baselines in 23 of 24 trunk–benchmark comparisons and are competitive with or superior to the GhostFaceNet++ variants, with the largest gains on the pose-sensitive evaluations. When trained with ArcFace on UMDFaces, FD-GhostFaceNet raises the best reported CP-LFW accuracy from 84.65% to 86.12% and CFP-FP from 87.60% to 89.56%; when trained on CASIA-WebFace, it improves the best reported CFP-FP from 90.10% to 90.94%. These results confirm that frequency decoupling strikes a favorable balance between compactness and accuracy, making both models strong candidates for deployment on resource-constrained devices.