Arabic Race Classification of Face Images

The Human face can tell us a lot about a person, their age, gender, expression and feelings, and may be even their race. There is really no substitute for face-to-face social interaction. In this paper, we develop a framework to classify human race from facial images into three categories, (Arabic, Asian, and Caucasian). This is done by transforming the face images to a lower dimension using Principal Component Analysis forwarded to a classifier constructer, using feed foreword neural network (scaled conjugate gradient) with different parameters such as different inputs (PCA components 40, 50, and 60), layers, neurons and training algorithms. The framework achieved high correct classification rate about 83.5 in certain scenarios. The experiments were conducted on a collection of images from FERET database. 160 images for the Asian category (2 images for one subject) and 160 images for the Caucasian category (2 images for one subject), and because there isn’t an Arabic category face image database we constructed our own, 120 images (2 images for one subject). The scaled conjugate gradient achieved about 15% higher than standard back propagation when tested on the conducted experiments. From the experiments the best result was achieved, i.e., in terms of the correct classification rate, time, and epochs needed to train network when the input to the network (PCA components) was 50 components. The framework provided evidence that the Arabic race exists. However, to confirm and further corroborate the findings, the framework needs to be applied on a larger sample.