Face Detection And Segmentation With Generalized Intersection Over Union Based On Mask R-Cnn

As a research hotspot of computer vision and information security, face detection has been widely developed in the past few decades. However, most of the existing detection methods only realize the location of the bounding box, which leads to background noise in the face features as well as limited accuracy of detection. To overcome these drawbacks, a face detection and segmentation method with Generalized Intersection over Union (GIoU) based on Mask R-CNN is proposed in this paper, which is called G-Mask. In this method, ResNet-101 is used to extract features, RPN is used to generate RoIs, and RolAlign faithfully retains the exact spatial locations to generate binary mask through Fully Convolution Network. In particular, to achieve better performance in multi-scale face detection tasks, we utilize GIoU as the bounding box loss function. Furthermore, a new face dataset with segmentation annotation information is constructed in this paper to train the model. The experimental results of the well-known benchmark FDDB and AFW show that the proposed G-Mask method achieves promising face detection performance compared with Faster R-CNN and the original Mask R-CNN method, and also can realize the instance-level face information segmentation while detecting.