协同表示下显著特征块筛选的人脸表情识别

Discriminative Patches Selection Based on Collaborative Representation for Facial Expression Recognition

针对人脸不同区域对于各种表情具有不同程度的区分性这一发现,提出一种基于协同表示(Collaborative Representation,CR)筛选特征块的人脸表情识别新方法.首先,通过协同表示学习训练样本,筛选出使得认证样本集中各类表情识别率达到最高的若干候选特征块;之后,在测试阶段,针对每个测试样本从候选块中自动筛选出独立的样本特征块,用于对该测试样本进行分类.与以往的块筛选方法不同,本文针对单个测试样本筛选出区分性的块.本文方法在CK+和JAFFE人脸表情库上的表现超越了其他特征块相关方法,并在不同分辨率和多种强度表情下取得了较好的识别效果.

Since the regions of face have different degrees of contributions to expression recognition, we proposed a new discriminative patches selection algorithm based on the Collaboration Representation （ CR ） for facial expression recognition. Firstly, candidate patches which classify validation samples well for each expression is selected by using the Collaborative Representation based Classification. Then,sample dependent patches which from candidate patch set are automatically further selected for different testing samples to classify each sample. Unlike to the previous patches selection method, we select discriminative patches for each testing sample. The pro- posed method performs well on CK ＋ and JAFFE facial expression databases compared to other relative feature patches method, and gets better accurancy under different resolutions and multi-intensity expression conditions.