WOMAN HIJAB DETECTION USING TRANSFER LEARNING

Abstract

Person clothing is considered one of many important issues related to Islamic rulings (Sharia). One of these issues is the woman dress where an Islamic woman is required to keep consistent wearing of Hijab (veils) when she is outside.  Based on Sharia and most of Islamic scholars, Hijab must cover woman’s hair, ears, and neck along with the top of chest. Classifying woman images into one who wearing Hijab or not can be significantly facilitated using the current approaches of deep learning. In this paper we proposed a structured method consisting of multiple steps for accurately classifying women’s images into Hijab and Non-Hijab by utilizing transfer learning approach of deep neural networks. We initially created a balanced and labeled dataset that includes 12,000 images from multiple sources, one half of the dataset for women wearing Hijab and the other for women not wearing Hijab. The dataset is then preprocessed for normalization techniques. After that we used three well-known pretrained models of transfer learning which are VGG16, Xception and MobileNetV2 to conduct our experiments. The feature extraction and fine-tuning strategies were used for examining the models. The selected models gave better performances when applying fine-tuning strategy where the accuracies values of 96.85%, 97.6% and 96.3% were achieved for VGG16, Xception and MobileNetV2 respectively. Our results proved the capability of transfer learning in detecting Hijab in women’s images in order to help individuals, institutions and others who are interested in Islamic dress.