遮挡自适应权重的人脸特征点定位算法

Face Landmark Localization with Adaptive Occlusion Weight

针对遮挡造成人脸定位失败和人脸特征点的遮挡标签匮乏等问题,提出了一种遮挡自适应权重的人脸特征点定位算法.基于HRNet,设计了一个遮挡自适权重损失函数,使遮挡点获得一个较小的权重用来减轻遮挡对特征提取的影响.在网络输出阶段,添加遮挡预测模块以获取特征点的遮挡度,对遮挡度进行线性变换作为热图回归任务的自适应权重.同时,通过在原图上生成随机大小、形状、颜色、纹理、透明度的遮挡及对应标签,进行数据集扩增.此外,根据预测坐标生成人脸的点特征图、边特征图、区域特征图以及切割图,将其与原图像融合后再输入主干网络,获得更好的人脸特征.本算法在COFW和300W等相关数据集上进行评估,取得了较好的准确性.

Occlusion is complex and random,which will break the structure of the face,resulting in the failure of facial landmark localization.At same time,there is a lack of occlusion labels for feature points.To solve these problems,we propose a algorithm for facial landmark localization with adaptive occlusion weight named occHRNet(occlusion_high-resolution network).Based on HRNet,an occlusion-adaptive weight loss function is designed so that the occlusion points get a smaller weight to reduce the impact of occlusion.In the network output stage,the occlusion prediction module is added to get the degree of occluded points.A linear transformation on the occlusion is perform to get the adaptive weight of the heat map regression task.Meanwhile,the dataset is amplified by generating occlusion of random size,shape,color,texture,transparency on the original image and corresponding labels.In addition,the point feature map,edge feature map,area feature map and cut map of the face are generated according to the predicted coordinates.Then,the multiple feature maps are fused with the original image before being fed into the backbone network to obtain better face features.The algorithm is evaluated on the COFW,300W nd other related datasets,and achieves better performance.