Research On Canonical Correlation Analysis Based On Different Multi-view Data Scenario With Applications | Posted on:2014-12-25 | Degree:Doctor | Type:Dissertation | Country:China | Candidate:X D Zhou | Full Text:PDF | GTID:1268330422979747 | Subject:Computer application technology | Abstract/Summary: | PDF Full Text Request | High-dimensional co-occurring data associated with an object frequently and abundantly emergein the real world. For example, an internet web page as an object can be represented as (co-occurring)page text and links to the page; a human can be represented as co-occurring visual and audio contents;an image can be represented by different kinds of features, such as color, texture and shape. This kindof data is usually called multi-view data or multi-represented data. Studies have shown thatmulti-view learning can acquire better performance than that of single-view learning through mutualpromotion between different views. Analyzing high-dimensional multi-view data to acquire anddiscover useful information and knowledge has become a hot research topic of pattern recognition,machine learning and data mining recently. Among these works, canonical correlation analysis (CCA)is one of the most widely adopted methods. Based on CCA, the major contributions of this thesis aresummarized as follows:(1) Since CCA does not utilize any class information at all and the combined features are theinput to the classifier in performing classification task based on CCA, we propose a new superviseddimension reduction method named Combined-Feature-Discriminability Enhanced CanonicalCorrelation Analysis (CECCA) for fully-supervised and fully-paired multi-view data. CECCA isdeveloped through incorporating discriminant analysis into CCA. Consequently, it optimizes thecombined feature correlation and discriminability simultaneously and thus makes the extractedfeatures more suitable for classification. In addition, since the most frequently used featurecombination methods are parallel combination and serial combination, we correspondingly proposetwo specific algorithms: CECCA_P (CECCA_Parallelism) and CECCA_S (CECCA_Serialization).(2) Based on the unsupervised and semi-paired multi-view data, we propose a novel CCA’ssemi-paired variant named Neighborhood Correlation Analysis (NeCA) to overcome the shortcomingof CCA, which is generally prone to overfitting and thus performs poorly, since its definition itselfmakes it only able to utilize those paired samples. NeCA is developed through incorporatingbetween-view neighborhood relationships into CCA, which uses unpaired samples in the multi-viewlearning way. In order to further improve the performance of NeCA, two extensions of NeCA aredeveloped: Laplacian-Regularization of NeCA (LRNeCA) and PCA-Regularization of NeCA(PRNeCA), which can utilize the unpaired samples in the single-view and multi-view learning way simultaneously. Promising experimental results on several popular multi-view data sets show itsfeasibility and effectiveness.(3) Faced up with semi-supervised and semi-paired scenario for low-resolution (LR) andhigh-resolution (HR) face training samples, a semi-paired and semi-supervised algorithm for LR facerecognition is developed. For the sake of utilizing HR and LR face training samples distributed indifferent feature space and limited labeled samples more effectively. The implementation of thealgorithm is divided into two stages. One stage is semi-paired learning and the other stage issemi-supervised learning. Promising experiments results on the Yaleã€AR and Extended Yale B facedatabases show the feasibility and effectiveness of the proposed method.(4) Under unsupervised and fully-paired scenario for LR and HR face multi-view data, amethod named Spatial Regularization of Canonical Correlation Analysis (SRCCA) is developed forLR face recognition, which can improve the performance of CCA by the regularization utilizingspatial information of different resolution faces. Since SRCCA has two spatial regularization termscorresponding to LR and HR respectively, different from single-resolution face recognition researchworks that only have one spatial regularization term. Furthermore, we empirically study and analysisthe impaction of LR and HR spatial regularization terms respectively on LR face recognition. | Keywords/Search Tags: | Dimensionality Reduction, Multi-view Learning, Single-view Data, Semi-paired Learning, Canonical correlation Analysis, Low-resolution face recognition, Semi-supervised Learning | PDF Full Text Request | Related items |
| |
|