An Efficient Face-based Age Group Detector on a CPU using Two Perspective Convolution with Attention Modules

Adri Priadana, Muhamad Dwisnanto Putro, Xuan-Thuy Vo, Kang-Hyun Jo

Abstract: Age detection has become incredibly substantial in various scenarios such as video surveillance, forensic applications, and advertising platform. An age detector is expected to operate on low-cost devices or CPU devices to minimize the budget of the implementation system. This work presents an efficient face based age group detector (Age-CPU) that can operate fluidly on a CPU. It proposes two perspectives convolution architecture with depthwise global attention modules (2PDG) on this detector. It applies two kernel sizes to consider different sizes of the feature area of the object reinforced with enhancing block. The depthwise layer on the attention module helps the architecture extract features more focused and deeply. It convolves each channel with an individual depthwise kernel. The architecture is trained and validated on the UTKFace and FG-NET datasets. 2PDG acquires competitive accuracy compared to other competitors’ architectures on the datasets. Furthermore, the proposed detector can operate 100 frames per second on a CPU device, which is speedy to execute in real-time.