Research On Partial Multi-label Learning Algorithm With Application To Image Semantic Understanding

In the traditional multi-label learning algorithm,each instance in the data set used for training is accurately labeled with multiple relevant labels.However,due to the rapid growth of data in recent years,accurate data sets are difficult to obtain in reality.The existing multi-label learning algorithms only considered the condition missing labels in data sets.However,in most cases,an instance in the data set is roughly assigned a group of labels,in which comprises some irrelevant noisy labels except for the relevant labels.In order to solve the problem of label redundancy in multi-label learning,a new learning framework,partial multi-label learning framework,was proposed in recent years.In the framework of partial multi-label learning,each instance contains a set of candidate label set,and the candidate label set contains all relevant labels and some noisy labels.The number of relevant labels is unknown but there is at least one.The task of partial multilabel learning is to obtain an accurate label prediction model by using redundant label data set.Partial multi-label learning difficulties mainly in two aspects.Firstly,because the training data sets contain noisy labels,the influence of noisy labels on the process of training need to be considered.Secondly,the features of the instance usually contain noise and also have impacts on the prediction model in the training process,like the influence of occlusion,blur and illumination in the image data set.Based on the above two points,this paper proposes two partial multi-label learning algorithms and carries out practical applications in image semantic understanding.A semantic understanding algorithm for partial multi-label image via multi-subspace representation.In the training process,the influence of redundant candidate label sets and image noisy features in the training process is considered.In order to reduce the negative effect of noisy labels on the prediction model,the initial noisy label space is decomposed into a low-dimensional label subspace and a label correlation matrix.Then the correlation between the instance features is used to map the initial noisy feature space to a lowdimensional feature subspace to reduce the influence of the noisy features on the prediction model.Finally,a Laplacian regularization term is introduced to constrain the label subspace to maintain the intrinsic structure consistency with the feature,and the orthogonality constraint is applied to the correlation between the features to ensure the discriminability of the feature subspace.Extensive experiments on different data sets prove the superiority of this method in partial multiple label learning and image semantic understanding.A semantic understanding algorithm for multi-view partial multi-label image via shared feature.In the training process,the effect of redundant labels in the training set and missing features of some views in the multi-views on the model training process is fully considered.Firstly,the missing view feature indicator matrix is used to decompose each view feature matrix and a shared view matrix is obtained to train the model.Then the initial label matrix with noise is factorize into a low-rank accurate label matrix and a sparse noise label matrix.The accurate label matrix is used to train the model for solving the influence of noisy labels.Finally,the Laplacian regularization term is introduced to constrain the low-rank accurate label matrix and the view matrix for ensuring the intrinsic structure consistency between the features and the labels.Besides,the low rank constraint is introduced to the coefficient matrix.Extensive experiments on different data sets are implemented which prove that this method has advantages in multi-view partial multilabel learning and image semantic understanding.