基于迁移学习和特征融合的人脸识别算法的研究

发布时间：2018-07-24 19:24

【摘要】：目前,传统的身份识别方法存在不便于携带、易丢失、易损坏和易被破解或窃取的风险。因此人脸识别得到了广泛的关注,由于其具有较强的稳固性、隐蔽以及个体间差别性,安全性得到了保障,应用领域也愈加广泛,例如安全、民用、军用等领域。人脸识别的研究对于其他图像处理领域发展同样有极其重大的意义。在实际应用情况下,人脸识别通常遇到小样本数据集问题,即所拥有的训练样本数目远小于待测人脸样本的大小,小样本数据集的问题会使得传统的特征提取方法和分类识别方法在人脸识别上很难获得较强的鲁棒性和较好的识别率。本文围绕"如何从人脸小样本数据集中有效地提取人脸特征和高效的实现人脸分类,从而提高人脸识别的鲁棒性和人脸识别的准确率"这一主线,研究和实现了基于迁移学习和特征融合的人脸识别算法,主要工作和创新点有:(1)针对标签类间样本具有相似性、同一标签类内样本具有差异性等问题,提出一种基于稀疏表示的迁移学习算法。本方法将稀疏表示和迁移学习相结合,首先待测人脸样本通过所有原始训练样本的线性表示,计算得到该待测人脸样本在原始训练样本上的稀疏系数矩阵。其次,定义类样本包括类测试样本和类训练样本,以实现类训练样本标签类内完全相同而标签类间完全不同,以及类测试样本到各标签类的距离相同的特性。将之前学习到的稀疏系数矩阵迁移到类训练样本上,以生成重构的类训练样本,从而利用了隐藏在不同稀疏系数向量中的判别性信息。重构的类训练样本通过稀疏表示模型得到类测试样本的重构误差,进而判定类测试样本的所属标签类。(2)针对分类可信度低等问题,通过融合原始样本和类样本的稀疏表示结果,提出一种基于权重融合方案的分类算法。原始人脸训练样本通过稀疏表示模型计算出待测人脸图像的重构误差,而重构的类训练样本也利用稀疏表示模型计算出类测试样本的重构误差,经过试验验证,两者重构误差之间的平均皮尔逊相关性较低,表明原始训练样本和重构的类训练样本对于同一个标签类的表示上具有互补性,因此对两者的重构误差赋予一定的权重并进行融合,利用融合后的结果来判定待测人脸样本的标签类,以获得更高的识别准确率。(3)为了克服稀疏表示算法仅考虑样本的稀疏性而忽略了样本内部的空间结构信息的缺点,提出了一种基于特征融合的人脸识别模型。本方法将稀疏表示模型的样本间稀疏性特征与均值哈希算法的样本内结构性特征相融合,通过稀疏表示模型计算得到待测人脸样本在原始人脸训练样本上的稀疏系数矩阵,同时计算出所有原始人脸训练样本和待测人脸样本的均值哈希特征,最终利用训练样本的均值哈希特征和稀疏系数矩阵来进行样本重构,待测人脸样本均值哈希特征减去重构样本,得到的值用来判定待测人脸样本的最终标签类,在稀疏表示的基础上提高了整体类别样本集合的识别准确度和鲁棒性。
[Abstract]:At present, the traditional identification method is not easy to carry, easily lost, easy to be damaged and easy to be cracked or stolen. Therefore, face recognition has been widely concerned. Because of its strong stability, concealment and individual differences, security has been guaranteed, and the application field is also widely used, such as security, civilian, military and so on. The research of face recognition is of great significance for the development of other image processing fields. In practical application, face recognition usually meets the problem of small sample data sets, that is, the number of training samples is far less than the size of the sample, and the problem of small sample data sets will make the traditional feature extraction side. The method and classification recognition method are difficult to obtain strong robustness and better recognition rate on face recognition. This paper focuses on "how to extract face features effectively from small sample data set and implement face classification efficiently, so as to improve the robustness of face recognition and the accuracy rate of face recognition". The main work and innovation points are as follows: (1) a migration learning algorithm based on sparse representation is proposed for the similarity of the samples between the labels and the differences in the sample of the same label class. This method combines sparse representation and migration learning to measure face samples first. Through the linear representation of all the original training samples, the sparse coefficient matrix of the tested face sample on the original training sample is obtained. Secondly, the class samples include class test samples and class training samples to achieve exactly the same classes in the class training sample labels and completely different among the label classes, as well as the class test samples to each label class. The sparse coefficient matrix previously learned is migrated to the class training sample to generate the reconstructed class training sample, thus using the discriminative information hidden in the different sparse coefficient vectors. The reconstructed class training samples get the reconfiguration error of the class test samples by the sparse representation model, and then determine the class test. 2. (2) in order to solve the problem of low classification reliability, a classification algorithm based on the weight fusion scheme is proposed by combining the sparse representation results of the original sample and the class sample. The original face training sample calculates the reconstruction error of the face image image by the sparse representation model, and the reconstructed class training sample is also beneficial. The sparse representation model is used to calculate the reconfiguration error of the class test sample. The experiment shows that the average Pearson correlation between the two reconstruction errors is low, which indicates that the original training samples and the reconstructed class training samples are complementary to the same label class. In order to overcome the shortcoming of the spatial structure information in the sample, a feature fusion based face recognition model is proposed to overcome the sparse representation of the sparse representation algorithm. (3) a face recognition model based on feature fusion is proposed. The sparsity feature of the sample model is fused with the structural features of the mean hash algorithm, and the sparse representation model is used to calculate the sparse coefficient matrix of the face sample on the original face training sample. At the same time, the mean hash characteristics of all the original face training samples and the face samples are calculated. The mean hash feature of the training sample and the sparse coefficient matrix are used to reconstruct the sample. The reconstructed sample is subtracted from the mean hash feature of the face sample, and the value is used to determine the final label class of the face samples to be measured, and the recognition accuracy and robustness of the whole class sample set is improved on the basis of the sparse representation.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】