摘要
针对现实环境中姿势、光照、表情及场景较大变化严重影响户外人脸识别算法识别性能的问题,提出了一种学习原型超平面(PHL)融合线性判别边信息(SILD)算法。首先,利用支持向量机(SVM)将弱标记数据集中的每个样本表示为一个原型超平面中层特征,使用学习组合系数从未标记的通用数据集中选择支持向量稀疏集;然后,在SVM模型组合稀疏系数的约束条件下,借助于Fisher线性判别准则最大化未标记数据集的判别能力,并使用迭代优化算法求解目标函数;最后,利用SILD进行特征提取,用余弦相似性度量完成最终的人脸识别。在户外标记人脸(LFW)和YouTube两大通用人脸数据集上,对PHL+SILD方法和低层特征+SILD方法在强度、LBP、Gabor特征和Block Gabor特征,平均精度、曲线下方面积(AUC)和等差率(EER)进行了比较。实验验证了所提算法的有效性及可靠性。
Factors will seriously impact recognition performance of wild face recognition algorithms, such as large variation of poses, illustration, expression and scene of real-world, a fusion face recognition algorithm based on Prototype Hyperplane Learning （PI-IL） and Side-Information based on Linear Discriminant （SILD） was proposed. Firstly, each sample in weak labeled data set act as a mid-level feature of prototype hyperplanes with Support Vector Machine （SVM） model, and SVM sparse set was selected from unlabeled generic data set by learning combination coefficient. Secondly, Fisher discriminative criterion was used to maximize discriminat ability under the constraint of combination sparse coefficients of SVM model, and the objective function was solved by iterative optimization algorithm. Finally, SILD was used to extract features and cosine similarity measure was used for face perception. PHL ＋ SILD method was compared with low-level feature ＋ SILD method in some indicators which is strength, LBP, Gabor feature and Block Gabor feature, average precision, AUC, EER. Experiments verify the effectiveness and reliability of the proposed algorithm.
出处
《计算机应用》
CSCD
北大核心
2014年第7期2044-2049,共6页
journal of Computer Applications
基金
湖南省教育厅科学研究项目(13C1070)
关键词
户外人脸识别
原型超平面学习
中层特征表示
支持向量机
线性判别边信息
wild lace recognition
Prototype Hyperplane Learning.（ I-＇HL）
mid-level teature representation
bupportVector Machine （SVM）
Side-Information based on Linear Discriminant （SILD）
