Semi-Dense U-Net: A Novel U-Net Architecture for Face Detection

Abstract

Face detection and localization has been a major field of study in facial analysis and computer vision. Several convolutional neural network-based architectures have been proposed in the literature such as cascaded approach, single-stage and two-stage architectures. Using image segmentation based technique for object/face detection and recognition have been an alternative approach recently being employed. In this paper, we propose detection of faces by using U-net segmentation architectures. Motivated from DenseNet, a variant of U-net, called Semi-Dense U-Net, is designed in order to improve the binary masks generated by the segmentation model and further post-processed to detect faces. The proposed U-Net model have been trained and tested on FDDB, Wider face and Open Image dataset and compared with state-of-the-art algorithms. We could successfully achieve dice coefficient of 95.68% and average precision of 91.60% on a set of test data from OpenImage dataset.