Algorithm Research On Discriminative Sparse And Low-Rank Representation

Image classification is a basic research task in the field of computer vision and pattern recognition.It is the foundation of many high-level vision tasks and plays a very critical role in the field of computer vision.The main task of image classification is to distinguish image categories based on the semantic information of the image.The discriminative sparse low-rank model is one of the important branches in image classification.After the researcher's work and experience accumulation,many research results have been obtained.However,facing challenges of real-world image classification tasks,how to efficiently obtain discriminative data representation is still a topic worthy of attention.Based on the discriminative sparse and low-rank model,this paper analyzes the shortcomings of the existing image classification models and designs new image classification models.The main work of this paper is as follows:(1)A low-rank feature-based double transformation matrix learning image classification algorithm is proposed.The original image contains a lot of complex information,and the existing least squares regression models may ignore the attention to the pressure of a single transformation matrix.Using the latent low-rank feature extraction algorithm to decompose the original data and eliminate the influence of sparse noise.Then introducing two transformation matrices to process the main features and salient features separately.Through these two measures,the classification pressure of a single transformation matrix in the traditional case is shared.The purpose of designing the double transformation matrices is to complete the relaxation of the transformation matrix,and to obtain a more flexible projection for the regression task.(2)An image classification algorithm based on the intra-class low-rank subspace is proposed.For the purpose of integrating feature extraction and regression tasks,subspace learning is used to learn an intermediate feature between the original data and the regression vector.On the one hand,the introduction of subspace projection and label space projection,has more flexibility than a single label space projection.On the other hand,the intra-class low-rank constraint imposed on the subspace makes the subspace feature a bridge between the original data and the label matrix,thereby enhances the overall performance of the model.At the same time,row sparsity helps the model to pay attention to the features that are intra-class low-rank in the data information,so that the model can obtain the required intermediate features.(3)A dictionary pair learning algorithm based on a structured classifier is proposed.Exploring and analyzing the naturally existing structured characteristics of the label matrix,using the extended label matrix to obtain a square matrix as a classifier,and combining with orthogonal constraints,a reversible classifier matrix is obtained.The obtained reversible classifier makes a structured constraint on the representation coefficients,so this reversible classifier is called a structured classifier.The structured classifier can transfer the structured characteristics of the label matrix to the representation coefficients explicably,and can also be used to obtain the label regression vectors for classification.Combining the existing research work of sparse representation and the dictionary pair learning,a new discriminative dictionary pair learning model is obtained.