基于自适应特征的遮挡人脸特征点定位算法

Facial Landmark Localization under Occlusion Based on Adaptive Feature

为了使人脸特征点定位算法在人脸被物体遮挡的情况下仍能快速、准确地检测特征点的位置,提出了一种基于自适应特征的遮挡人脸特征点定位算法。该方法首先检测每个特征点的遮挡状态,即先训练一个逻辑回归模型,通过所有特征点周围的纹理特征快速地估计出每个特征点被遮挡的概率值;然后根据每一个特征点被遮挡的概率自适应地调整该特征点纹理特征的权重,使得被遮挡概率较大的特征点获得较小的权重值,减小人脸遮挡对特征的影响,提高特征点定位的准确度。实验结果表明,本文算法的特征点定位的平均误差达到5.94%,遮挡检测准确率/召回率达到80%/72.84%。

In order to make the facial landmark localization algorithm detect the position of feature points quickly and accurately when the face is occluded by objects, this paper proposes a facial landmark localization under occlusion method based on adaptive feature extraction. The method first detected the occlusion state of each feature point, that is, first trained a logistic regression model, and quickly estimated the probability value of each feature point being occluded with the texture features around all landmarks. Then, adaptive weights were assigned to each feature point according to their estimated occlusion probability, so that the feature points with larger occlusion probabilities were assigned smaller weight values, therefore the impact of facial occlusion on the feature was decreased, and the accuracy of landmark localization was improved. Quantitative experiments on the challenging COFW benchmark show that the proposed method obtains the state-of-art results in terms of localization accuracy and occlusion detection, with an average localization error of 5.94% and a precision/recall of 80%/72.84%.