基于多视图特征投影与合成解析字典学习的图像分类

Multi-view feature projection and synthesis-analysis dictionary learning for image classification

针对目前存在的合成解析字典学习方法不能有效地消除同类样本之间的差异性和忽略了不同特征对分类的不同影响的问题,提出了一种基于多视图特征投影与合成解析字典学习(MFPSDL)的图像分类方法。首先,在合成解析字典学习过程中为每种特征学习不同的特征投影矩阵,减小了类内样本间的差异对识别带来的影响;其次,对合成解析字典添加鉴别性的约束,使得同类样本具有相似的稀疏表示系数;最后通过为不同类型的特征学习权重,充分地融合多种特征。在公开人脸数据库(LFW)和手写体识别数据库(MNIST)上进行多项对比实验,MFPSDL方法在LFW和MNIST数据库上的训练时间分别为61.236 s和52.281 s,MFPSDL方法相比Fisher鉴别字典学习(FDDL)、类别一致的K奇异值分解(LC-KSVD)、字典对学习(DPL)等字典学习方法,在LFW和MNIST上的识别率提高了至少2.15和2.08个百分点。实验结果表明,所提方法在保证较低的时间复杂度的同时,获得了更好的识别效果,适用于图像分类。

Concerning the problem that the existing synthesis-analysis dictionary learning method can not effectively eliminate the differences between the samples of the same class and ignore the different effects of different features on the classification, an image classification method based on Multi-view Feature Projection and Synthesis-analysis Dictionary Learning （MFPSDL） was put forward. Firstly, different feature projection matrices were learned for different features in the process of synthesis-analysis dictionary learning, so the influence of the within-class differences on recognition was reduced. Secondly, discriminant constraint was added to the synthesis-analysis dictionary, so that similar sparse representation coefficients were obtained for samples of the same class. Finally, by learning different weights for different features, multiple features could be fully integrated. Several experiments were carried out on the Labeled Faces in the Wild （LFW） and Modified National Institute of Standards and Technology （MNIST） database, the training time of MFPSDL method on LFW and MNIST databases were 61. 236 s and 52. 281 ,~. Compared with Fisher Discrimination Dictionary Learning （ FDDL）, Lable Consistent K Singular Value Decomposition （LC-KSVD） and Dictionary Pair Learning （ DPL）, the recognition rate of MFPSDL method on LFW and MNIST was increased by at least 2.15 and 2.08 percentage points. The experimental results show that MFPSDL method can obtain higher recognition rate while keeping low time complexity, and it is suitable for image classification.