Research On Image Classification Algorithms Based On Linear Regression And Dictionary Learning

With the development of machine learning and deep learning,image classification has become one of the most popular research areas in recent years.Especially,face recognition has arisen more and more attention from experts and scholars due to its wide range of application scenarios.The core idea of image classification is to enable the computer to extract effective visual feature information from pictures and predict its specific categories precisely,which is also a key and difficult problem in the field of image classification.Based on linear regression and dictionary learning,this paper aims to improve the recognition accuracy of noisy pictures and enhance the discriminativeness of the recognition model.A series of classification algorithms are proposed and applied to multiple image classification tasks.Large experimental results have proved that the three algorithms proposed improved the performance of the image classification system.Specifically,the main research content of this thesis is introduced as follows.For the problem of image classification with noise,a linear regression algorithm based on negative label relaxation technology is proposed.Based on linear regression,we use negative label relaxation technology to reduce the class margins between different classes to drop the impact of noise.In addition,we also exploit manifold learning,and locality preserving projections,according to the principle that samples from the same class should be kept close after they are transformed into a new space,we design a suitable regularized item to prevent the problem of over-fitting.We added a L2-norm constraint on the loss function,and conduct experiments with other classification algorithms on multiple public face data sets for a comparsion.The experimental results show the feasibility and effectiveness of this algorithm.In order to extend the above model from face recognition to other areas of image classification,a linear regression algorithm based on L2,1-norm constraint is proposed.The algorithm is improved on the basis of the linear regression algorithm based on negative label relaxation technology,replacing the L2-norm constraint on the loss function with the L2,1-norm constraint,while the regularization term remains unchanged.The rotation invariance of the 2,1-norm is used to impose row sparsity constraints on deep features,so as to get key features from samples.This not only improves the accuracy,but also obtains satisfactory results in the scene recognition data set.For the problem of insufficient discrimination in dictionary learning,an elastic-net based discriminative K-SVD dictionary learning algorithm is proposed.The algorithm adds a classification error term on the objective function based on K-SVD,and then the elastic-net is put on the sparse coefficient matrix to reduce the correlation coefficient between each atom in the dictionary.Hence,the dictionary learned has a stronger representation ability and a higher accuracy simultaneously.We conducted experiments on different face data sets and object data sets.Compared with other algorithms,this algorithm has a stonger discriminative ability.