Publication: Lightweight face recognition utilizing convfacenext model and triplet branch attention module
Loading...
Date
2023-09-01
Authors
Hoo Seng Chun
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
There are several motivations to develop lightweight model with low
parameters and FLOPs that include the feasibility to be deployed in edge devices,
which require small model size and low power consumption. ConvNeXt block that is
taken as a reference has high parameters and FLOPs. Thus, an Enhanced ConvNeXt
(ECN) block is developed in the first part of this research. The ECN block deploys
small kernel size to reduce the parameters and FLOPs. Several ECN blocks are
stacked to develop the ConvFaceNeXt models. UMD Faces with ground truth labels
is used as the training dataset for ConvFaceNeXt, using ArcFace loss function. For
testing, seven image-based datasets (i.e., LFW, CALFW, CPLFW, CFP-FF, CFP-FP,
AgeDB-30 and VGG2-FP) and two template-based datasets (i.e., IJB-B and IJB-C)
are used. During testing, ConvFaceNeXt is only used to obtain the face embedding
while discarding the ArcFace loss function. Although the computation has been
decreased, the performance is not compromised. Next, an attention module with both
channel and spatial branches, which is Lightweight Triplet Branch Attention Module
(LTBAM) has been developed as the second part of this research. Concretely, each
of these branches encodes only necessary information through one convolution
operation using small kernel size to reduce the parameters and FLOPs. Since
LTBAM needs to be integrated with ConvFaceNeXt, the training and testing process
is the same as ConvFaceNeXt. As opposed to Dimension-Aware Attention (DAA)
module, although LTBAM reduces the parameters and FLOPs by 4.52% and 0.23%
respectively, the average accuracy is increased by 0.65% for the seven image-based datasets as well as 0.92% and 0.57% for the corresponding IJB-B dataset and IJB-C
dataset. In conclusion, both ConvFaceNeXt and LTBAM not only reduce the
parameters and FLOPS compared to the respective previous works but also improve
on the verification performance.