Automatic Gender Classification through Face Segmentation

Khalil Khan,Muhammad Attique,Ikram Syed,Asma Gul

doi:10.3390/sym11060770

Abstract

Automatic gender classification is challenging due to large variations of face images, particularly in the un-constrained scenarios. In this paper, we propose a framework which first segments a face image into face parts, and then performs automatic gender classification. We trained a Conditional Random Fields (CRFs) based segmentation model through manually labeled face images. The CRFs based model is used to segment a face image into six different classes—mouth, hair, eyes, nose, skin, and back. The probabilistic classification strategy (PCS) is used, and probability maps are created for all six classes. We use the probability maps as gender descriptors and trained a Random Decision Forest (RDF) classifier, which classifies the face images as either male or female. The performance of the proposed framework is assessed on four publicly available datasets, namely Adience, LFW, FERET, and FEI, with results outperforming state-of-the-art (SOA).

Highlights

Automatic gender classification is a fundamental task in computer vision, which has recently attracted immense attention
The performance of points localization models is significantly affected in cases such as background, occlusion, face image rotation, and if the face images are of poor quality
Superpixel algorithms aim to over-segment the face image by grouping pixels that belong to the same object

Summary

Introduction

Automatic gender classification is a fundamental task in computer vision, which has recently attracted immense attention. Gender classification is still an arduous task due to various changes in visual angles, face expressions, pose, age, background, and face image appearance. It is more challenging in the un-constrained imaging conditions. Face parts are located through prior facial landmarks localization models [1,2,3,4], and gender classification is performed. The performance of points localization models is significantly affected in cases such as background, occlusion, face image rotation, and if the face images are of poor quality. Localization models are facing severe problems, and we approached the gender classification task in an entirely different way. For more detailed information about gender classification methods, a review paper in [15] can be explored

Methods

Results

Conclusion