Face Detection with Improved Faster-R-CNN
Meng Wang, Fen Zheng, Jiangdong Lu · 2022
With the continuous development of convolutional neural network, object detection technology has made great progress. As a representative of the two-stage object detection algorithm, faster-r-cnn achieves high precision in object detection tasks. However, there are still many difficulties and challenges using fater-r-cnn to detect faces. For example, tiny faces in the images are not detected, and side faces or occluded faces are prone to be detected as false detections. In response to those questions raised above, this paper improves faster-r-cnn algorithm in the following aspects: we replaced the feature extraction network in the faster-r-cnn algorithm and used feature pyramid networks, which is more beneficial to detect tiny faces. Besides, we used classification loss function which is more conducive to classification task, and for the special sizes of faces, we set the anchor ratio matching mechanism. In addition, we used suitable activation function to increase the nonlinear fitting ability of the whole network, and for the problem of the training set of WIDER FACE database, we cleaned it in order to get better experimental results. As a consequence, our method achieved 96.2% mAP on FDDB and 93.6%, 88.5%, 62.6% mAP respectively on the easy, medium, hard set of WIDER FACE, and our method could run at 10 FPS on Nvidia GTX 1080Ti GPU.