Using estimated arithmetic means of accuracies to select features for face-based gender classification

Selecting the appropriate features is essential in building a good classifier. This paper aims to use the approach of estimating the arithmetic means of accuracies (ameans) in selecting the features used in a face-based gender classification. In a face-based gender classification, there are many pixels of the input image that may not aid the classification process, such as those belonging to the background. The experiments show that this approach outperforms the approach based on mean difference especially on the data having relatively high variance by up to 2.14%. Compared to the classifier which does not use any feature selection approach, implementing the feature selection approach based on ameans estimation in a gender classification problem increases the accuracy by up to 7.86%. The experiments also show that the face-based gender classifications rely on the presence of long hair on subjects in the images to make their decision.