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

With the exponential growth of the massive images and multimedia social data in recent years,multi-label learning is widely used in text categorization,image recognition and annotation,multimedia data analysis,biomedical diagnostic and personalized recommendations due to its multi-semantic representation,which can effectively meet the demands of different users in complex scenarios.Existing multi-label learning algorithms are mainly faced with four challenges:the number of training samples are usually huge while the number of labeled data is limited;insufficient utilization of correlation between different labels;the tags annotated on the training images are unclean or incomplete;the extracted low-level input features fail to explain the high-level semantic label concepts.In response to the aforementioned problems,this paper proposes two multi-label learning methods and applies the later approach to the filed of image semantic understanding.A semi-supervised dual low-rank feature mapping algorithm for multi-label image annotation is presented.Firstly,it introduces a linear self-recovery model to repair the noisy or missing items in original label matrix innovatively,which can maintain the low-rank structure of the label space.In order to capture the label correlation in abundant training samples,the dual low-rank trace norm regularization items of feature mapping matrix and coefficient matrix of self-recovery model are introduced,and they can control the model complexity as well.Besides,graph Laplacian regularization is exploited to maintain the local manifold structure by making use of the amounts of unlabeled images by enforcing the local geometric structure on both labeled and unlabeled instances.Extensive experiments demonstrate the effectiveness and efficiency of the proposed framework.It will be more powerful especially when the training samples are insufficient and the label information is noisy.A patch-based latent variable model of multi-label learning algorithm is proposed.Most multi-label learning approaches may lead to "semantic gap" when there is complex semantic concepts of high-level in training images.Being aware of this,the paper introduces a patch-based latent variable model based on statistic and probability to re-represent the original features of input images.In other words,the patch-based feature vectors of input images are low-level features;the feature matrix of latent visual entities is the re-representation of the input feature,which can be called mid-level output;the output label space is the collection of high-level semantic concepts.The introduction of the latent variable model can not only capture the implicit relations between the patches globally,but also contribute to the feature dimension reduction and the decrease of the cost of time and memory.And the summarization of the latent representation vectors and the Laplacian regularization across patches facilitate to utilize the local information from geometric structure.A large number of experiments show the superior annotated performance of the proposed framework in image semantic understanding.Furthermore,a simple image annotation system is implemented based on Matlab platform,which can effectively predict the high-level complex semantic labels for scene images.