Sparse Representation in Classification

Sparse representation with learning-based overcomplete dictionaries has recently achieved impressive results in signal representation and outperforms traditional methods in tasks such as signal denosing and image inpainting. Inference algorithms also benefit from sparse representation because of its effectiveness in removing data corruption such as dense noise and partial occlusion and the ability to model the high-level factors underlying the data. The representation, also known as feature extraction has contributed to machine learning algorithms with good performance. In this thesis, the problem of combining sparse representation-based methods and discrimination-based methods in inference tasks to gain benefit from both types of approaches has been studied. Specifically, three algorithms were developed to address problems in three different applications. Sparse representation was firstly employed as an interpretable method to model pathological human gait in order to help diagnosis. The used sparse representation algorithms simultaneously train linear classifiers during the representation process. Then the problem of inter-class coherence in multi-class data is addressed to encourage classes to be sparsely represented in a more discriminative way. A framework was proposed to separate the representative and discriminative components though sparse representation to enlarge the disparity of learned class models. Finally, support vector machine and a special form of sparse representation, the archetype analysis, is combined in a latent model with temporal constraints to extract ordinal key poses from 3D skeleton videos for action recognition. Experiments using well-known datasets were designed to evaluate the proposed algorithms and demonstrate their effectiveness.