基于字典扩展的稀疏表示鲁棒人脸识别算法研究

发布时间：2019-01-08 09:46

【摘要】：人脸识别以其友好性和便捷性,成为生物识别的研究热点。基于稀疏表示人脸识别方法因诸多优点受到广泛关注,本文基于国内外相关学者研究基础上,分成三章节对稀疏表示人脸识别算法进行改进。首先,针对训练样本字典学习仅包含全局信息、缺乏局部信息的不足,引入与类别相关的原子字典,提出基于原子与分子字典联合扩展的加权稀疏表示人脸识别方法。通过对各类训练样本进行主成分分析,构造主成分基原子字典,同时将训练样本字典作为分子字典。进而,利用原子字典与分子字典结合得到扩展字典模型。测试时,根据测试样本与扩展字典基之间的距离进行加权得到与当前测试样本关联的重构字典集,最后对测试样本稀疏重构,利用残差进行分类判别。其次,传统稀疏表示分类字典模型在稀疏重构或重构残差时只考虑了辨别性信息,识别性能受到限制的问题,因此提出一种新的字典构造方法。对所有训练样本进行主成分分析,得到共享性分子字典,将训练样本作为辨别性原子字典,从而构造为扩展字典。然后,使用扩展字典对测试样本进行稀疏表示。最后,计算每类样本的表示概率,选取最大概率类别进行分类。最后,针对训练样本和测试样本均存在光照及遮挡时,破坏图像低秩结构问题,提出基于监督低秩子空间恢复的正则鲁棒稀疏表示人脸识别算法。首先,将所有训练样本构造成矩阵,并进行监督低秩矩阵分解,分解为低秩类相关结构,低秩类内差异结构和稀疏误差结构,利用主成分分析找到类相关结构低秩子空间的变换矩阵,再通过变换矩阵将训练样本和测试样本投影到低秩子空间。最后,通过正则鲁棒稀疏编码进行加权分类识别。
[Abstract]:Because of its friendliness and convenience, face recognition has become the research hotspot of biometric recognition. Face recognition based on sparse representation has attracted much attention because of its many advantages. Based on the research of relevant scholars at home and abroad, this paper divides into three chapters to improve the sparse representation face recognition algorithm. Firstly, because the training sample dictionary learning contains only global information and lacks local information, an atomic dictionary related to category is introduced, and a weighted sparse representation face recognition method based on the joint expansion of atomic and molecular dictionaries is proposed. Through the principal component analysis of all kinds of training samples, the principal component based atomic dictionary is constructed. At the same time, the training sample dictionary is used as a molecular dictionary. Furthermore, an extended dictionary model is obtained by combining atomic dictionary with molecular dictionary. According to the distance between the test sample and the extended dictionary base, the reconstructed dictionary set associated with the current test sample is obtained. Finally, the sparse reconstruction of the test sample is performed, and the residual error is used to classify and discriminate. Secondly, the traditional sparse representation classification dictionary model only considers the discriminative information and the recognition performance is limited in sparse reconstruction or reconstruction residuals, so a new dictionary construction method is proposed. All training samples are analyzed by principal component analysis and shared molecular dictionaries are obtained. The training samples are used as discriminative atomic dictionaries and thus are constructed as extended dictionaries. Then, the test samples are represented sparsely using an extended dictionary. Finally, the representation probability of each kind of sample is calculated, and the maximum probability category is selected for classification. Finally, a regular-robust sparse representation face recognition algorithm based on supervised low-rank subspace restoration is proposed to destroy the low-rank image structure when both training and test samples have illumination and occlusion. First of all, all the training samples are constructed into matrices, and the supervised low rank matrix is decomposed into low rank correlation structure, low rank class difference structure and sparse error structure. The transformation matrix of low rank subspace with class correlation structure is found by principal component analysis, and then the training sample and test sample are projected to low rank subspace through the transformation matrix. Finally, weighted classification and recognition are carried out by regular robust sparse coding.
【学位授予单位】：燕山大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】