New Data Representation Method Combined With Low-rank Representation And Dictionary Learning

With the rapid development of information technology,data plays an increasingly important role in computer sciences such as computer vision,classification,and clustering.Huge and high-dimensional data impose a burden on the algorithm and affect the performance of the algorithm.Exploring efficient data representation has always been an important issue in computer science.Dictionary learning is an effective feature learning method that can learn the sparse representation of data.Self-representation,initially expressed in locally linear embedding,is now widely used for low-rank representation.Dictionary learning and low-rank representation have been widely used in the fields of pattern recognition and computer vision.There are more and more works to combine these two.Therefore,it makes sense to study the algorithm of combining dictionary learning and low-rank representation.This paper combines dictionary learning and low-rank representation algorithms to seek a better data representation model.The contributions of this thesis are listed as following:1? Aiming at the characteristics of the self-representation model that can capture the construction of the datasets,the properties of the LRR model are discussed in detail in this paper.However,in the actual application fields,the dictionary is utilized to represent the data rather than the data itself.In this paper,we have proposed a more general data representation model with dictionary learning properties and derived a closed-form solution by our rigorous mathematical derivation.Compared with the iterative solution,the convergence time required for our method is significantly shortened.The model with dictionary property we proposed making the low-rank self-representation model a more general method while retaining the low-rank property.2? A novel data representation learning framework,which combines the dictionary learning and self-representation,is proposed.In this model,we obtain a new data representation combined with sparse dictionary learning and low-rank representation,which further mining the structure information of the data by two-way complementation.Our model not only supplements the low-rank structural information but also satisfies the sparse characteristics.Our model introduces low-rank and sparsity into the same model framework.The low-rank representation learned by the model promotes the learning of the dictionary,and the dictionary,in turn,improves the learned low-rank representation,enabling the model to obtain the sparse representation and the low-rank representation at the same time.3? The model proposed in this paper can be applied in two applications.First,compared with the dictionary learned by the classical dictionary learning algorithm,the dictionary learned by our model has fewer meaningless atoms and more atoms that are representative.The other application is used for extracting data representation.Because the model is unsupervised learning,the data labels are not required for supervising and constraining.Therefore,the model can be applied to extract representation before any classification and clustering algorithms.Experiments in this paper proved that the model could obtain better results on classification and clustering tasks.