Research On Feature Learning And Classification Methods Based On Semi-supervised

The problem of image feature learning and classification are hot issues in the field of computer vision and pattern recognition,and have been widely concerned by academia and industry for many years.With the rapid development of computer and internet of things technology,a large number of image data resources are constantly produced,and the categories and volumes are also growing rapidly.Researchers are free to access these image data and use these to achieve the purpose of image feature learning and classification model training.However,due to the expensive labeling cost,only part of the acquired image data has labels,and the unlabeled data cannot be used for the training of traditional supervised learning models.In order to solve this problem,the semi-supervised learning method came into being.The basic idea is to use both labeled and unlabeled data in the process of model training.While expanding the training data,it uses pseudo-label prediction and data correlation mining technologies,iteratively train the model to effectively improve the learning ability of the model.This paper takes the semi-supervised learning as the basic research object,combined with efficient data representation models to explore its specific application methods in image feature learning and classification problems.The aim is,with only partially labeled image data,to solve the problem of robust feature subspace learning,feature selection and classification prediction,and improve the performance of existing learning and classification models.The main research contents of this paper include the following aspects:(1)We propose a low-rank constraint discriminative feature learning method based on semi-supervised.The method uses non-negative low rank representation coefficient to constrain characteristic subspace learning which improves the robustness of the method.At the same time,the non-negative low rank representation coefficient is used to construct the Graph Laplace matrix for label learning.In addition,the learned label is combined with the labeled data,and a label regression term is constructed to improve the discriminability and applicability of the method.Then,an iterative algorithm of convergent objective function is designed,and the effectiveness of the method is demonstrated in the original dataset and the noisy dataset respectively.The experimental results show that the proposed method has outstanding performance on the original dataset and show better robustness on the noisy dataset.(2)We propose a robust feature selection method based on semi-supervised.The method uses a scaling factor to rescale the regression coefficient in the least square method to order the importance of all features of the sample and explains it in theory.At the same time,thel_2,1 norm constraint is applied to the projection matrix,which is helpful for the model to learn the sparse solution and improve the performance of feature selection.In addition,this method can effectively remove the noise information of the original data in reality by using the low-rank constraint on the representation coefficient.Then,an effective objective function solution algorithm is designed and its convergence is guaranteed.By verifying the method on multiple datasets,it can be seen that the method shows stable performance on both the original dataset and the noisy dataset.(3)We propose a semi-supervised feature learning method based on bi-graph regularized low-rank representation.In order to retain the local structure information in the data,this method combines LatLRR model with graph learning to carry out a graph similarity constraint on two low-rank representation coefficients in LatLRR respectively.In addition,the representation coefficient for the projection space in LatLRR is equivalent to the characteristic projection matrix of the sample,and then a ridge regression model is constructed by using the projected samples.Therefore,a representation coefficient for the projected sample can be learned without prior sample error.Furthermore,the representation coefficients of the original samples are combined with those of the ridge regression model to construct a graph Laplace matrix with useful label learning.Corresponding algorithms are designed for each step of the proposed method,and this method is validated on multiple datasets.It can be seen from the existing experimental results that compared with the common semi-supervised methods based on low-rank representation and graph-based learning,the performance of the proposed method is more outstanding.