联合判别性低秩类字典与稀疏误差字典学习的人脸识别

发布时间：2018-03-21 05:15

本文选题：低秩类字典　切入点：稀疏误差字典　出处：《中国图象图形学报》2017年09期 　论文类型：期刊论文

【摘要】：目的由于受到光照变化、表情变化以及遮挡的影响,使得采集的不同人的人脸图像具有相似性,从而给人脸识别带来巨大的挑战,如果每一类人有足够多的训练样本,利用基于稀疏表示的分类算法(SRC)就能够取得很好地识别效果。然而,实际应用中往往无法得到尺寸大以及足够多的人脸图像作为训练样本。为了解决上述问题,根据基于稀疏表示理论,提出了一种基于联合判别性低秩类字典以及稀疏误差字典的人脸识别算法。每一类的低秩字典捕捉这类的判别性特征,稀疏误差字典反映了类变化,比如光照、表情变化。方法首先利用低秩分解理论得到初始化的低秩字典以及稀疏字典,然后结合低秩分解和结构不相干的理论,训练出判别性低秩类字典和稀疏误差字典,并把它们联合起来作为测试时所用的字典;本文的方法去除了训练样本的噪声,并在此基础上增加了低秩字典之间的不相关性,能够提高的低秩字典的判别性。再运用l1范数法(同伦法)求得稀疏系数,并根据重构误差进行分类。结果针对Extended Yale B库和AR库进行了实验。为了减少算法执行时间,对于训练样本利用随机矩阵进行降维。本文算法在Extended Yale B库的504维每类32样本训练的识别结果为96.9%。在无遮挡的540维每类4样本训练的AR库的实验结果为83.3%,1 760维的结果为87.6%。有遮挡的540维每类8样本训练的AR库的结果为94.1%,1 760维的结果为94.8%。实验结果表明,本文算法的结果比SRC、DKSVD(Discriminative K-SVD)、LRSI(Low rank matrix decomposition with structural incoherence)、LRSE+SC(Low rank and sparse error matrix+sparse coding)这4种算法中识别率最高的算法还要好,特别在训练样本比较少的情况下。结论本文所提出的人脸识别算法具有一定的鲁棒性和有效性,尤其在训练样本较少以及干扰较大的情况下,能够取得很好地识别效果,适合在实际中进行应用。
[Abstract]:Objective because of the influence of illumination change, facial expression change and occlusion, the human face images collected from different people are similar, which brings a great challenge to face recognition, if there are enough training samples for each kind of people. The classification algorithm based on sparse representation can achieve good recognition results. However, in practical applications, face images with large size and enough face images can not be used as training samples. Based on sparse representation theory, a face recognition algorithm based on joint discriminant low rank dictionaries and sparse error dictionaries is proposed. Methods the initialized low rank dictionary and sparse dictionary are obtained by using the theory of low rank decomposition, and then the discriminant low rank dictionary and sparse error dictionary are trained by combining the theory of low rank decomposition and structural incoherence. The method in this paper removes the noise of the training sample and increases the non-correlation between the low-rank dictionaries. The sparse coefficient is obtained by using l 1 norm method (homotopy method) and classified according to the reconstruction error. Results the experiments are carried out on Extended Yale B library and AR library. In order to reduce the execution time of the algorithm, For training samples, random matrix is used for dimensionality reduction. The recognition result of 32 samples per class of 504 dimension in Extended Yale B library is 96.9. The experimental result of AR library with 4 samples training in 540D without occlusion is 83.3 dimensional. The result is 87.6. The AR library with 540 dimensions of occlusion and 8 kinds of samples trained is 94.1D and 1,760D, 94.80.The experimental results show that, The result of this algorithm is better than that of the four algorithms (SRC DKSVD discriminative K-SVD and LRSI low rank matrix decomposition with structural incorencein LRSE SC(Low rank and sparse error matrix sparse). Conclusion the face recognition algorithm proposed in this paper is robust and effective, especially in the case of less training samples and more interference. Suitable for practical application.
【作者单位】：南京航空航天大学自动化学院;
【基金】：国家自然科学基金项目(61473148,U1531110) 江苏省普通高校专业学位研究生创新计划(SJLX16_0106)~~
【分类号】：TP391.41

【相似文献】