多级调优的人脸检测网络

Face detection network based on multi-stage optimization

为了提高人脸检测的速度和精度,提出了一个端到端的人脸检测网络。该网络包含两个模块:粗筛模块和精调模块。粗筛模块负责过滤背景候选框,仅保留少量高质量候选框,以提高检测的速度;精调模块则负责进一步遴选剩余的候选框并对其进行调整来适应真实的人脸,以提高检测的精度。另外,选择性地对一些层的特征进行融合从而进一步提高网络的精度。在FDDB数据集上,该网络在500次假阳性误报的情况下取得了95.71%的离散得分和75.29%的连续得分。

In order to improve the speed and accuracy of face detection, an end-to-end face detection network was proposed. It includes two main modules: coarse-scale module and fine-scale module. The coarse-scale module was used to filter the background proposals and keep only a small amount of high-quality proposals to speed up the detection. The fine-scale module was used to further filter the remaining proposals and adjust the selected proposals to improve the detection accuracy. Besides, features of some layers were selected to fuse together in order to further improve sccuracy of the network. On FDDB dataset, the proposed network achieves 95.71% of discrete score and 75.29% of continuous score with 500 false positives.