基于级联网络和残差特征的人脸特征点定位

为进一步提高人脸特征点定位精度，探究当前广泛用于人脸关键点定位的全卷积神经网络（FCN）架构的原理和缺陷，讨论FCN核函数在特征点定位中引入的副作用，即训练和测试时评判准则不一致的问题. 理论分析该问题存在的可能性和普遍性，设计实验验证在实际场景下此问题存在的广泛性. 提出结合残差特征的沙漏网络结构并将其应用于人脸特征点检测；提出多级沙漏网络的级联结构，并将其与经典的栈式沙漏网络进行对比分析. 实验结果表明：二级级联结构获得了与四级栈式结构相当的特征点定位精度，大幅降低了模型参数量和时间复杂度. 所提方法在300-W数据库的困难子集上的平均归一化误差为6.84%，优于已有最好方法.

Facial landmark localization based on cascaded hourglass network with residual features

The principles and defects of full convolutional network (FCN), which was widely utilized in facial landmark localization, were studied to improve the facial landmark localization accuracy. Discuss the side effects introduced by the kernel function in the feature of FCN, that the evaluation criteria were inconsistent during training and testing. Firstly, theoretically analyze the possibility and the universality of this problem, and then design experiments to verify the existence of this problem in actual situation. To solve this problem, a hourglass network structure was proposed for facial landmark localization combining residual features; the cascaded hourglass network structure was given. The experimental results show that the two-stage cascade structure can obtain comparable accuracy compared with the four-stage stack structure, which means that the model parameter quantity and time complexity will be reduced greatly. The average normalization error of the proposed method on the difficult subset of the 300-W database was 6.84%, which is better than the previous best result.