基于稀疏表示分类的人脸识别

发布时间：2018-02-06 06:00

本文关键词： 人脸识别稀疏表示 Fisher判别加权迭代最小二乘　出处：《大连海事大学》2017年硕士论文　论文类型：学位论文

【摘要】：基于稀疏表示的人脸识别是一种高效的人脸识别算法,它假设一幅测试图像可以在训练图像上稀疏表示,然后计算最小的类误差来分类。稀疏表示分类对遮挡、光照和噪声具有鲁棒性,而且有很好的实验效果。但是它直接使用训练样本作为字典,导致了它不能有效的表示测试图像;用原始样本作为字典也很难充分利用隐藏在训练样本间的信息。基于稀疏表示的人脸识别的编码方法一般采用的是用l1范数近似l0范数,但是这个方法也有缺点。一方面,在某些程度上,这些方法的收敛速度比较慢。另一方面,当测试样本处于非常大的随机遮挡时,稀疏解通常非常稠密以致于很难找到正确的测试样本的类别。针对上述提出的问题,本文从两个方面改进了算法。第一个改进算法是对字典学习进行改进,在已有算法的基础上提出了简化的Fisher判别字典学习的简化算法,通过合并区分保真项中的协同惩罚项和判别惩罚来简化模型,有效地提高了人脸识别的效率。第二个改进算法是对稀疏编码进行改进,在迭代重加权最小二乘法的基础上提出了加权的迭代重加权最小二乘算法,并与其它求解l1范数最小化和l0范数最小化的算法进行了比较。在AR、ORL和YALE人脸数据库上的实验结果表明,本文的两种算法比已有算法具有更好的重建效果和识别效果。
[Abstract]:Face recognition based on sparse representation is an efficient face recognition algorithm, which assumes that a test image can be represented sparsely on the training image, and then the minimum class error is calculated to classify. Illumination and noise are robust and have good experimental results, but it directly uses training samples as dictionaries, which leads to its inability to represent test images effectively. It is difficult to make full use of the information hidden between the training samples by using the original sample as the dictionary. The coding method based on sparse representation is generally using l 1 norm to approximate l 0 norm. On the one hand, the convergence rate of these methods is slow to some extent. On the other hand, when the test samples are in a very large random occlusion. Sparse solutions are usually so dense that it is difficult to find the correct types of test samples. This paper improves the algorithm from two aspects. The first one is to improve dictionary learning. Based on the existing algorithms, a simplified Fisher discriminant dictionary learning algorithm is proposed, which simplifies the model by combining the cooperative penalty terms and discriminant penalties in fidelity items. The second improved algorithm is to improve the sparse coding. Based on the iterative reweighted least square algorithm, a weighted iterative reweighted least squares algorithm is proposed. It is compared with other algorithms for minimizing l 1 norm and l 0 norm. The experimental results on ARDORL and YALE face database show that the proposed algorithm is effective. The two algorithms in this paper have better reconstruction effect and recognition effect than the existing algorithms.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】