IRNet: An Improved RetinaNet Model for Face Detection

The current face detection methods mainly focus on overlaying network layers to improve the detection accuracy. However, in practical applications, these huge models cannot achieve the real-time detection. In order to solve the aforementioned problems, an improved RetinaNet (IRNet) model for face detection is proposed. In the field of face detection, feature fusion module is a common method to solve multi-scale detection problems. In this paper, a new feature fusion module N-FPN is introduced. The N-FPN module has been proven to improve detection accuracy. We design experiments to select the appropriate weight decay parameters. The IRNet finally achieves 91.77%, 89.26%, and 76.59% detection results on Easy, Medium, and Hard datasets of Wider Face, respectively. In addition, the IRNet obtained better detection results than RetinaNet on the FDDB dataset.