级联神经网络人脸关键点定位研究

Research on face alignment based on cascade convolutional neural networks

人脸关键点定位是计算机视觉的一部分,在人脸识别、人脸表情识别、人脸动作捕捉等工作中有重要的作用.非约束条件下人脸关键点定位,其难点在于人脸关键点位置在复杂环境下呈现非线性变化,影响人脸关键点定位的精准性.现提出基于级联卷积神经网络的人脸关键点定位方法,分析了级联深度模型全局回归阶段多尺度特征融合对人脸关键点定位的影响;同时提出了一种具有可学习参数的人脸关键点定位损失函数.经过大量实验表明,这里提出的人脸关键点定位算法能够有效的提高针对非约束条件下人脸关键点定位精确度.

Face alignment is meaningful in computer vision which is an important procedure in face recognition, emotional analysis, and motion capture, etc. Face alignment on unconstrained conditions is difficult clue to the differences between face appearances and shapes. We proposed an improved cascade convolutional neural network and added multi scale feature fusion to enhance the ability of network feature learning in the global regression stage. A novel loss function was proposed too. The experimental results show that the proposed network effectively improves the accuracy of the model trained on unconstrained conditions.