An Efficient Face Gender Detector on a CPU with Multi-Perspective Convolution

Intelligent digital signage demands software to work lightly in real-time on embedded devices. Smart digital signage requires gender detection to deliver relevant and targeted ads automatically. This study presents an efficient real-time face gender detector (FGenCPU) that can run smoothly on a CPU implemented in digital signage to decide immediately and provide relevant and targeted ads while the audience faces it. The proposed architecture contains Multi-Perspective Convolution (MPConvNet) consisting of two main modules: backbone and recognition. An efficient extractor module with a multi-perspective contextual track through the various kernel sizes is utilized to apprehend different feature areas of the object. This architecture employs few filters in every convolution layer and generates few parameters. It makes the detector operate at high speed for real-time. The training is conducted on the UTKFace dataset to generate efficiently weighted models. The MPConvNet achieves competitive performance with other common and light architectures on UTKFace, Adience, and LFW datasets. Furthermore, the detector can run 38 frames per second when performing on a CPU in real-time.