Person's discriminating visual features for recognising gender: LASSO regression model and feature analysis

Abstract

Gender is one of the demographic attributes of a person, which is considered as a soft trait in the area of biometric. Several studies have been conducted to extract gender information based on a person's face image, gait pattern, fingerprint, iris, speech and hand geometry. In this paper, we concentrate on predicting gender using a person's image aesthetic, which has never been studied before. We propose a visual preference model for discriminating males from females using LASSO regression. The preference model uses 57 dimensional feature vector containing 14 different perceptual image features. The model is evaluated on a database of 34,000 images from 170 Flickr users (110 males and 60 females). Results show that maximum and average accuracy of predicting gender are around 91.67% and 84.38%, respectively, on 100 random sampling of training and testing datasets. The proposed method outperforms all existing state-of-the-art methods. In this paper, we also address two important research questions: which features are impacting the discrimination of male-female visual preferences and how many images are sufficient for predicting a person's gender.