基于改进SBR算法的人脸特征点稳定检测

基于图像的特征点检测器在静态图像上取得了卓越的性能，然而这些方法应用于视频或序列图像时其精度和稳定性显著降低。配准监督(Supervision-by-Registration, SBR)算法利用光流算法(Lucas-Kanade, LK)追踪，可通过无标注视频训练针对视频的特征点检测器，已取得较好的结果，但LK算法仍存在一定局限性，导致检测的特征点序列在时空上的连贯性不强。为获得精准、稳定、连贯的人脸特征点序列检测效果，提出了平滑一致性损失函数、权重掩码函数对传统SBR网络模型进行改进。网络中添加长短期记忆网络(Long Short-Term Memory, LSTM)提高模型训练鲁棒性，在模型训练中使用平滑一致性损失函数提供稳定性约束，获得准确且稳定的人脸视频特征点检测器。在300VW、Youtube Celebrities数据集上的验证显示，SBR改进模型将人脸视频特征点检测的标准化平均误差(Normalized Mean Error, NME)从4.74降低至4.56,且视觉上人脸特征点检测的抖动显著减少。

Facial feature point stability detection based on improved SBR algorithm

The feature point detector based on image achieves excellent performance on static image.However,when these methods are applied to video or image sequence,their accuracy and stability are significantly reduced.The Supervision-by-Registration(SBR) algorithm uses Lucas-Kanade(LK) algorithm to track,which can train the feature point detector for the video through unlabeled video and has achieved good results.However,the LK algorithm still has some limitations,leading to the weak coherence of the feature point sequence detected in space and time.In order to obtain accurate,stable and coherent result of facial feature point sequence detection,smoothing consistency loss function and weight mask function are proposed to improve the traditional SBR network model.Long Short-Term Memory(LSTM) network is added to the network to improve the robustness of model training,and smoothing consistency loss function is used to provide stability constraints in model training to obtain accurate and stable facial video feature point detectors.The experiment results on 300VW and Youtube Celebrities datasets show that the improved SBR model reduces the normalized mean error(NME) of face video feature point detection from 4.74 to 4.56,and the jitter of facial feature point detection in visual sense is significantly reduced.