摘要
针对传统基于回归的人脸对齐算法在人脸尺度归一化时会造成纹理的损失,以及为了提升算法模型的泛化能力必须扩充数据集重新训练而导致训练时间增加,甚至出现不收敛、不可计算等问题,提出一种基于尺度自适应与增量式学习(IL)的人脸对齐方法来提高定位精度。首先,建立初始人脸形状与标准人脸形状的映射关系;然后,通过映射关系实现纹理特征在原图上的提取和人脸尺度的归一化;最后,利用算法模型在新的数据集上进行增量式的学习,快速提高原模型的泛化能力。实验结果表明,与传统回归方法相比,所提方法有更高的对齐精度,特别是在AFW数据集(68个特征点)上提高了2~4个百分点;在10万级别的大数据集(5个特征点)上,所提方法的鲁棒性比基于深度学习的方法高1~2个百分点。同时,所提的增量式学习方法不仅适用于人脸对齐场景下的回归模型求解,还适用于其他应用场景下回归模型的求解。
To solve the problems of traditional regression-based methods such as the loss of texture caused by face scale normalization, costing more time for retraining expanded data set to improve the generalization ability of the original model, and even potential non-convergence and incomputability, a face alignment method based on scale self-adaption and Incremental Learning （IL） was proposed. Firstly, the mapping relationship between the initial face points and the standard face points was established. Secondly, the extraction of the texture features on the original image and the normalization of the face scales were achieved via the mapping relationship. At last, incremental learning was applied to the new data set by using the existing models, which improved the generalization ability of the original model quickly. The experimental results show that the proposed method performs higher alignment accuracy than traditional regression-based methods. On the AFW （Annotated Faces in the Wild） dataset （68 feature points）, the accuracy is increased by 2 to 4 percentage points; and on a 100000-level large dataset （5 feature points）, the robustness is increased by 1 to 2 percentage points compared to the methods based on Deep Learning （DL）. In addition, the proposed method is not only suitable for face alignment regression model, but also applicable for solving other regression models.
作者
陈平
龚勋
CHEN Ping;GONG Xun(School of Information Science and Technology,Southwest Jiaotong University,Chengdu Sichuan 611756,China)
出处
《计算机应用》
CSCD
北大核心
2018年第7期2064-2069,共6页
journal of Computer Applications
基金
国家自然科学基金资助项目(61772435)~~
关键词
人脸对齐
增量式学习
尺度自适应
级联回归
鲁棒性
face alignment Incremental Learning （IL）
scale self-adaption
cascaded regression robustness
