A Convolutional Neural Network for Gender Recognition Optimizing the Accuracy/Speed Tradeoff

Antonio Greco,Vincenzo Vigilante,Mario Vento,Alessia Saggese

doi:10.1109/access.2020.3008793

Antonio Greco, Vincenzo Vigilante + Show 2 more

Open Access

https://doi.org/10.1109/access.2020.3008793

Copy DOI

Abstract

Gender recognition has been among the most investigated problems in the last years; although several contributions have been proposed, gender recognition in unconstrained environments is still a challenging problem and a definitive solution has not been found yet. Furthermore, Deep Convolutional Neural Networks (DCNNs) achieve very interesting performance, but they typically require a huge amount of computational resources (CPU, GPU, RAM, storage), that are not always available in real systems, due to their cost or to specific application constraints (when the application needs to be installed directly on board of low-power smart cameras, e.g. for digital signage). In the latest years the Machine Learning community developed an interest towards optimizing the efficiency of Deep Learning solutions, in order to make them portable and widespread. In this work we propose a compact DCNN architecture for Gender Recognition from face images that achieves approximately state of the art accuracy at a highly reduced computational cost (almost five times). We also perform a sensitivity analysis in order to show how some changes in the architecture of the network can influence the tradeoff between accuracy and speed. In addition, we compare our optimized architecture with popular efficient CNNs on various common benchmark dataset, widely adopted in the scientific community, namely LFW, MIVIA-Gender, IMDB-WIKI and Adience, demonstrating the effectiveness of the proposed solution.

Highlights

Gender recognition from faces is one of the basic capabilities of the human beings
We propose an optimal Deep Convolutional Neural Networks (DCNNs) architecture tuned for gender recognition
We first select a known architecture that leverages the latest devices from the state of the art of deep learning; we show different variants of the chosen architecture to study the effect of the variation on both classification accuracy and prediction latency

Summary

INTRODUCTION

Gender recognition from faces is one of the basic capabilities of the human beings. Extending this capability to machines is of great interest in many application areas. Greco et al.: Convolutional Neural Network for Gender Recognition Optimizing the Accuracy/Speed Tradeoff processing units, even being quite powerful, with 32 bit parallelism, vector co-processors and capabilities of floating point computation, are typically equipped with ARM processors and only low resources in terms of memory and storage This is true in applications like digital signage, in which a small smart camera, with around 512 MB of RAM and 16 MB for storing the whole application, including the model of the network, needs to perform the classification of several faces in real-time to quickly customize the promotional content. The main contributions of this paper are the following: 1) we demonstrate with a comprehensive experimental analysis that it is possible to preserve the gender recognition accuracy by carefully modifying the architecture of a CNN; 2) we propose a network architecture devised for gender recognition, optimized by reducing the input size, the number of feature maps and the number of layers of an existing network architecture, achieving a performance comparable with state of the art but can be suitably applied in embedded applications with real-time constraints

RELATED WORK

EXPERIMENTS

Findings

CONCLUSION