Research On Multi-label Learning With Missing Labels For Image Classification

Image classification is an important research hotspot in the field of machine learning and pattern recognition,whose purpose is to train a classifier based on image data sets with known and complete labels to predict the labels of new samples.Traditional image classification is usually a single-label classification problem,that is,one image corresponds to one label.However,images in the real world are often polysemic,they are often associated with multiple class labels.Aiming at the polysemy problem which is difficult to be solved by traditional single-label learning frameworks,multi-label learning frameworks have emerged at the historic moment.Multi-label learning explicitly assigns the subsets of the relevant labels to the sample objects.Most of the existing multi-label learning algorithms assume that the label set of the training samples is complete.However,this is not the case in the practical applications.In practice,the labels of training samples are usually manually labeled,which is timeconsuming and labor-intensive.Subjectivity of manually labeling and polysemy of image can easily lead to the problem of missing labels.In addition,with the development of the Internet and social media,a large amount of image data sets labeled and uploaded by users are easily available,but these image data sets obviously have the problem of missing labels.How to train a robust and effective classifier using these easily available data sets with missing labels is still an urgent problem.This thesis mainly studies the multi-label image classification with missing labels.The main research contents are as follows:1.Using the inherent correlation between labels and the similarity between examples,a multi-label image classification algorithm based on instance-wise and label-wise correlations is proposed.First,a linear reconstruction strategy is used to measure the similarity between each instance and its neighborhoods.Second,the low rank representation of the label matrix is utilized to explore the high-order correlations between the labels.Then,a weighted least square loss function is utilized to ensure the consistency between the given labels and the predicted labels.Finally,the Laplacian manifold regularization is utilized to combine the two correlations and the loss function to generate the final objective function.Experiments on multiple image datasets prove that the algorithm can effectively handle the problem of missing labels.2.Based on low rank feature mapping and low rank label recovery,a dual-low-rank multi-label image classification algorithm with missing labels is proposed.First,the algorithm explores the label correlations in feature space by assuming that the feature mapping coefficient matrix of the algorithm is low rank.Then,the framework explores the high-order correlations between labels by assuming the label correlation matrix to be a low rank matrix.And the low-rank label correlation is used to recover the missing labels of the original label matrix.Then,the extraction of label correlations,the recovery of labels,and the training of the model are combined by guaranteeing the consistency between the predicted labels and the recovered labels.Finally,the instance-wise correlation is introduced through regularization.