三角稀疏回归分类器及其在稳健人脸识别中的应用

发布时间：2018-05-28 16:52

本文选题：回归拟合 + 复数空间　；参考：《南京大学学报(自然科学)》2017年06期

【摘要】：以稀疏表示为代表的回归分类方法对于高斯噪声具有较好的鲁棒性,但容易受到训练样本中离群点数据的影响导致欠拟合或过拟合.通过探索余弦函数对离群数据呈现的周期不敏感特性,使用余弦函数来刻画回归残差,并在复数域空间进行稀疏回归,提出了三角稀疏回归分类器(TSRC)模型.考虑到模型的非凸特性,普通的迭代算法难以获得全局最优解.因此,通过三角函数演算与核函数技巧将TSRC转化为一个凸优化问题,使用交替方向乘子法(ADMM)对模型进行求解,核函数的计算过程从核空间角度解释了模型对离群点鲁棒的本质原因,通过欧拉公式能够完全避开复数域的计算过程,从而起到加速的作用.在AR,Extend-YaleB及NUST-RF带有遮挡和光照变化的人脸识别数据集上进行了识别率与速度的实验,验证了所提出模型的有效性,在Extend-YaleB数据集上测试了所提出方法在不同尺度的训练样本下的运行效率,并与现阶段先进方法进行了对比.
[Abstract]:The regression classification method, represented by sparse representation, is robust to Gao Si noise, but it is easy to be influenced by outlier data in training samples, resulting in under-fitting or over-fitting. By exploring the periodicity insensitivity of cosine function to outlier data, using cosine function to depict regression residuals and sparse regression in complex domain space, a triangular sparse regression classifier (TSRC) model is proposed. Considering the nonconvexity of the model, it is difficult to obtain the global optimal solution by the common iterative algorithm. Therefore, TSRC is transformed into a convex optimization problem by trigonometric function calculus and kernel function technique, and the model is solved by alternating direction multiplier method (ADMMM). The calculation process of kernel function explains the essential reason that the model is robust to outliers from the point of view of kernel space. Eulerian formula can completely avoid the calculation process of complex field and thus play an accelerating role. Experiments on the recognition rate and speed of the face recognition data set with occlusion and illumination variation in ARN Extend-YaleB and NUST-RF are carried out to verify the effectiveness of the proposed model. The efficiency of the proposed method in different scale training samples is tested on the Extend-YaleB dataset and compared with the advanced methods at present.
【作者单位】：金陵科技学院计算机工程学院;南京理工大学计算机科学与技术学院;东南大学自动化学院;
【基金】：江苏省高等学校自然科学研究重大资助经费项目(17KJA520001) 江苏省研究生科研创新计划项目(KYCX17_0361)
【分类号】：TP391.41

【相似文献】