A Facial Gender Detector on CPU using Multi-dilated Convolution with Attention Modules

Facial gender detectors have evolved into a vital component of an intelligent advertisement display platform. It is helpful to assist a decision of delivering appropriate advertisements to each audience. To reduce system costs, applications deployed in this platform must be able to run on a CPU. This work proposes a facial gender detector (FG-CPU) that can be implemented on a CPU device to support an advertising display platform. The proposed CNN model consists of a multi-dilated convolution with attention modules (MudaNet). The multi-dilated convolution is applied to capture multi-scale features in an efficient manner. The attention module is used to rectify the quality of the feature map. This work’s training and validation process is conducted on the UTKFace, the Labeled Faces in the Wild (LFW), and the Adience Benchmark datasets. As a result, the proposed CNN model is proven to compete with other common and lightweight competitors’ CNN models on these three datasets. Regarding speed, the detector can operate 49.19 frames per second in real-time on a CPU device.