An innovative deep learning approach for Arabic race recognition

Abstract

In computer vision, human race detection has become a critical application across many domains, such as security and customized marketing. Deep learning approaches, such as convolutional neural network (CNN), have played an essential role in improving human race detection. Nevertheless, detecting Arabic race is still a field that has received little attention. In this paper, an Arabic human race dataset comprising the following classes: Gulf, Levant, Sudan, Egypt, and North Africa (excluding Egypt) has been collected and proposed as a starting point for Arabic race classification. This dataset has been evaluated using a simple CNN-based model and other transfer learning models: DenseNet121, VGG16, and ResNet50. The difficulty in classifying these regions lies in the similarity of border areas in people’s features and in intermarriage between different regions, which helps transfer genetic traits that distinguish one region from another. The best results in recall, F1-score, precision, and accuracy were obtained by the DenseNet121 model, which achieved an average accuracy of 0.746 across five folds.