基于鲁棒损失函数的人脸关键点检测
首发时间:2017-08-24
摘要:人脸关键点检测是一种对人脸识别至关重要的计算机视觉任务。L2损失通常应用于该任务的卷积神经网络(CNN)优化,但它会受到训练过程中异常值的严重影响,异常值定义为该样本估计与平均训练样本估计有极大偏差的情况。为了解决这个问题,本文提出了一种使用Tukey鲁棒损失函数的回归CNN模型,该鲁棒损失是M-estimators之一且实现了对异常值的鲁棒性。本文的方法与传统的L2损失相比,不仅具有更好的收敛值,而且训练过程中收敛速度更快。该方法也在两个公开数据集中显示出比目前较先进的方法很有竞争力甚至更好的结果。
关键词: 鲁棒损失 人脸关键点检测 M-estimator CNN 异常值
For information in English, please click here
Facial Keypoints Detection with Robust Loss Function
Abstract:Facial keypoints detection is a computer vision task which is essential for face recognition. L2 loss is commonly applied in the Convolutional Neural Networks (CNN) optimization of this task which is seriously affected by outliers in the training process (outlier in this context is defined by a sample estimation is an extreme deviation from the mean training sample estimations).To solve this problem, this paper propose a regression CNN model that use Tukey's bisquare function, one of the M-estimators achieve robustness to outliers, as the loss function. Our method shows not only better convergence value but also faster convergence rate in training process compare to traditional L2 loss. Our approach also shows comparable or better results than current state-of-the-art approaches in two publicly available datasets.
Keywords: Facial Keypoints Convolutional Neural Network Robust M-estimator Outlier
论文图表:
引用
No.4740625862381150****
同行评议
共计0人参与
勘误表
基于鲁棒损失函数的人脸关键点检测
评论
全部评论0/1000