A Fast Real-time Face Gender Detector on CPU using Superficial Network with Attention Modules

A gender detector has become an essential part of digital signage to support the decision to provide relevant ads for each audience. Application installed in digital signage must be capable of running on low-cost or CPU devices to minimize system costs. This study proposed a fast face gender detector (Gender-CPU) that can sprint in real-time on CPU devices implemented on digital signage. The proposed architecture contains a superficial network with attention modules (SufiaNet). This architecture only consists of three convolution layers, making it super shallow and generating a small number of parameters. In order to redeem the lack of a super shallow network, the global attention module is assigned to improve the quality of the feature map resulting from the previous convolution layers. In the experiment, the training and validation process is conducted on the UTKFace, the Adience Benchmark, and the Labeled Faces in the Wild (LFW) datasets. The SufiaNet gains competitive accuracy compared to other common and light architectures on the three datasets. Moreover, the detector can run 84.97 frames per second on a CPU device, which is fast to run in real-time.