Abstract:Recent studies have shown that better recognition performance can be achieved by representing symmetric positive definite matrix as points on Riemannian manifolds in many computer vision tasks, which catches nonlinear structure information of dictionary atoms. However, most of the existing dictionary learning algorithms do not consider the locality information together with label information and pairwise positive constraint of dictionary atoms in the process of dictionary learning, which limits the recognition performance of the algorithms. Besides, most of the existing algorithms locally approximate Riemannian manifolds only from the tangent space, which cannot effectively describe the distribution of samples. Inspired by the kernel method, a new approach of incorporating Riemannian kernels into locality constraint, label-embedding and pairwise positive constraint dictionary learning algorithm is proposed and successfully applied to vision classification. Firstly, vision data are described by symmetric positive definite matrices, which are expressed as points on Riemannian manifolds. Secondly, with the help of the recently proposed Riemannian kernel, the graph Laplacian matrix of the learned dictionary is constructed on Riemannian manifolds rather than in the Euclidean space as the conventional algorithms do, which retains the locality information of the dictionary atoms on Riemannian manifolds. Following, label-embedding and pairwise positive constraint terms of dictionary atoms retaining the discriminative information of the learned dictionary are constructed on Riemannian manifolds, and thus, the sparse coding and dictionary learning mathematical model on Riemannian manifolds are established by the locality-constrained, label-embedding and pairwise positive constraint algorithm. Thirdly, a semi-supervised Riemannian kernel dictionary learning algorithm based on locality-constrained (LCSRKDL) is presented according to the convex optimization methods. Finally, an iterative updating algorithm is constructed to optimize the object function, and a linear classifier is used to complete the classification of test samples. The experimental results on five visual classification datasets show that the LCSRKDL algorithm has greatly improved the recognition accuracy.

Classification with Incoherent Kernel Dictionary Learning

A Brief Summary of Dictionary Learning Based Approach for Classification (revised)

A dictionary learning approach for classification: separating the particularity and the commonality

A Classification-Oriented Dictionary Learning Model: Explicitly Learning the Particularity and Commonality Across Categories.

Online Discriminative Dictionary Learning for Image Classification Based on Block-Coordinate Descent Method

Synthesis Linear Classifier Based Analysis Dictionary Learning for Pattern Classification

Class-Aware Analysis Dictionary Learning for Pattern Classification.

Synthesis K-SVD Based Analysis Dictionary Learning for Pattern Classification

Anomaly Detection with Selective Dictionary Learning

Image classification based on kernel dictionary learning

Reduced Kernel Dictionary Learning

Class-Oriented Discriminative Dictionary Learning for Image Classification

Unidirectional Representation-Based Efficient Dictionary Learning

Analysis Dictionary Learning: An Efficient and Discriminative Solution

Incoherent Dictionary Learning for Sparse Representation

Supervised Class-Specific Dictionary Learning for Image Classification.

Learning Dictionary on Manifolds for Image Classification

Projective Dictionary Pair Learning for Pattern Classification.

RETRACTED: A Semi-supervised Riemannian Kernel Dictionary Learning Algorithm Based on Locality-Constrained for Image Classification (retracted Article)

Support Discrimination Dictionary Learning For Image Classification

Discriminant sparse coding for image classification