CASPN:基于级联空间金字塔的人脸关键点定位网络

CASPN: cascaded spatial pyramid network for face landmark localization

针对非限制环境下人脸关键点定位的诸多干扰因素,如遮挡、阴影,以及如何设计更加轻量、快速的神经网络的问题,尝试并联不同空洞率的空洞卷积应用于人脸关键点定位,在保持特征分辨率的同时,快速增大并且获取多重感受野信息来获得更全局的语义信息,同时结合特征融合为精确定位关键点与关键点猜测提供丰富的上下文信息,以此提出一种实时、轻量级、高检测精度的人脸关键点定位网络。该网络的参数量约为2.7million,模型只有10.6 MB,在保持高检测精度的同时,在GTX1080设备上可达约150 fps的处理速度。目前在流行的数据集中也获得了优异的评估结果,其中在WFLW测试集中取得了5.40%的mean error与7.36%的failure rate。

Focused on the issue that faces exposed in unconstrained environments within occlusion,shadow etc,and considered the efficient and speed of a neural network,this paper proposed a real-time and lightweight convolutional neural network for face alignment,which employed paralleled atrous convolution to capture larger and multi-scale receptive fields and cascaded spatial pyramid networks within feature fusion for refinement and context information for keypoints speculation. The model can be merely 10. 6 MB and reach approximately 150 fps per face with high precision on GTX1080 device. Now,it achieves a good performance comparable to the state-of-the-art algorithms,5. 40% mean error and 7. 36% failure rate on WFLW testset.