Human Face Detector with Gender Identification by Split-Based Inception Block and Regulated Attention Module

Abstract

Smart digital advertising platforms have been widely arising. These platforms require a human face detector with gender identification to assist them in the determination of providing relevant advertisements. The detector is also prosecuted to identify the gender of a masked face in post-coronavirus situations and demanded to operate on a CPU device to lower system expenses. This work presents a lightweight Convolution Neural Network (CNN) architecture to build a gender identification integrated with face detection to respond to these issues. This work proposes a split-based inception block to efficiently extract features at various sizes by partially applying different convolution kernel sizes, levels, and regulated attention module to improve the quality of the feature map. It produces slight parameters that drive the architecture efficiency and can operate quickly in real-time. To validate the performance of the proposed architecture, UTKFace and Labeled Faces in the Wild (LFW) datasets, modified with an artificial mask, are utilized as training and validation datasets. This offered architecture is compared to different lightweight and deep architectures. Regarding the experiment results, the proposed architecture outperforms masked face gender identification on the two datasets. In addition, the proposed architecture, which integrates with face detection to become a human face detector with gender identification can run 135 frames per second in real-time on a CPU configuration.