Abstract: The region-based convolution network has been widely used in object detection, attracting extensive researcher’s interest. Aiming at the problem of face detection, this paper proposes an improved face detection algorithm based on Region-based Fully Convolutional Networks (R-FCN). In order to make the model training more complete, the online hard example mining method is used to relax the constraints of positive and negative samples, which extends the scope of the training set. For the overlapping problem of face targets, a linear non-maxima suppression method is adopted to avoid missing detection  of overlapping faces. The experimental results on the face detection database (FDDB) show that the improved R-FCN model has a higher accuracy than the original R-FCN model.

Key words: face detection, deep learning, object detection, fully convolutional networks