Collecting large training dataset of actual facial images from facebook for developing a weighted bagging gender classifier

Abstract

Many of previous gender classifiers have a common problem of low accuracy in classifying actual facial images taken in real environments since they were learned in restricted environments. Therefore, this study proposes to swiftly collect uncontrolled actual facial images from Facebook to construct training dataset and proposes a weighted bagging gender classifier which utilizes a Facebook dataset to increase the classification accuracy. In the proposed gender classification scheme, utilization of unique features extracted with the LBP (Local Binary Patterns), Gabor wavelets, and HOG (Histogram of Oriented Gradients) algorithms was proposed. Also, a weighted bagging classification scheme was proposed to vote on the final gender classification. The developed classifier showed comparatively high accuracy rate of 94.68% using the LFW (Labeled Faces in the Wild) dataset, a common evaluation test dataset for face recognition.

Bagging ensemble for deep learning based gender recognition using test-time augmentation on large-scale datasets

“…Earlier studies showed that LBP and Gabor jets offer better performance than raw pixels when used with SVM and variants. For this reason, these two features are extensively used in the literature [15,[26][27][28][29][30][31].…”

“…The bagging method for gender recognition is studied in [27,30]. In [27], multiple SVM based linear models are used to create bagging of classifiers and stacking for gender recognition.…”

“…They use histograms obtained from LBP and HOG descriptors. In [30], the weighted bagging method is used in combination with LBP histograms and Gabor wavelets. Using a dynamic weighting method they compared the majority voting with the weighted bagging model.…”