Framework Model of Facial Analysis for Gender Classification Using Convolutional Neural Network

Abstract

Classification is a technique used for solving problems. Several problems are solved with this technique. Gender classification is gaining ground due to different areas of applications such as surveillance, security, and monitory, etc. Different authors have presented different research articles in the domain of gender classification and adopted several methods for analysing facial images in other to predict or classify the images. These methods adopted are either traditional algorithms, hybridised techniques, or neural networks to obtain better accuracy and reliability. This article is aimed at developing a model where gender can be classified. Successful classification needs a robust method with good experimental analysis that is why we present a gender classification using a Convolutional Neural Network for reliability and accuracy using a local dataset. Although, most of the articles in this research area made use of popular datasets such as FERET, AT & T, FACE94, AR to mention but few and/or compare two or more datasets to know the one with the best performance accuracy. Our state of heart method was used on local data set where sizable numbers of images (490 images) were captured and five different augmentations such as blur, top hat, lightening, etc were carried out on the images. The dataset was divided into two with 70% of the images used for training and the remaining 30% for testing. This was done with the use of a random selection algorithm. Required 227by227by3 image size was pre-trained by AlexNet a CNN. The experimental results generated several tables, Area under Curve (AUC) and Confusion Matrix.  Our proposed ConvNet on our local dataset improves gender classification accuracy. In conclusion, the parameters for evaluation of performance were calculated and their Average performance scores were highlighted in bold. For Precision (89.6272); Recall (89.6276); Accuracy (92.8094) and F1-score (89.6237).  The best performance average score was 92.8094 under the Accuracy.