Factorizing and Reconstituting Large-Kernel MBConv for Lightweight Face Recognition

Abstract

In the past few years, Neural Architecture Search (NAS) has exhibited remarkable advances in terms of neural architecture design, especially on mobile devices. NAS normally use hand-craft MBConv as building block. However, they mainly searched for block-related hyperparameters, and the structure of MBConv itself was largely overlooked. This paper investigates that factorization and reconstitution can promote the efficiency of large-kernel MBConv and thus proposes FR-MBConv (Factorizing and Reconstituting large-kernel MBConv). Compared to large-kernel MBConv with the same receptive field, our FR-MBConv has fewer number of parameters and less computational cost, dramatically increased depth and nonlinearity. In addition, from the perspective of feature generation mechanism, FR-MBConv can be equivalent to more regular convolutions. We combine FR-MBConv with MobileNetV3 to build a lightweight face recognition model. Extensive experiments on face recognition benchmark demonstrate that our lightweight face recognition model outperforms the state-of-the-art lightweight model. Even on large scale face recognition benchmark IJB-B, IJB-C and MegaFace, our lightweight model also achieves comparable performance with large models.