Robust Semi-Supervised Multi-Label Learning By Triple Low-Rank Regularization With Application To Automatic Image Annotation

Annotating images automatically has been an active research topic due to its usage on web search.The advancement and growth of digital visualization has led to a vast number of images being available on the web and most of these images are without descriptions.Automatic image annotation is associated with multiple labels,and the image annotation is a typical multi-label classification problem,as a result multi-label learning-based method have attracted significant interests due to their perfect semantic representations for image annotation which assigns multiple labels to the images and also has good scalability with other machine learning paradigms.Despite its significance,multi-label learning based has its own limitations.The limitations are:training images may only have been annotated with a partial set of the label,the labels can be noisy or corrupted,due to its unavailability and quality of manual tags there is a limited number of labeled images and also correlations among different labels and integrate.The objective of this study is therefore,is to improve the annotation performance in order to address the limitations multi-label learning.Most studies,as regards to automatic image annotation application can effectively handle exploiting the semantic correlations among different labels using different machine learning techniques.However,most studies are not robust in the aspect of handling label noises or incomplete labels,and also when compared to vast number of unlabeled images there is limited number of labeled images.In this study,by addressing these limitations proposes a new multi-label learning predictive model named robust semi-supervised multi-label learning by triple low-rank regularization that helps improve the performance of the annotation.Specifically,the proposed model first utilizes the advantage of Low-Rank Representation(LRR)in feature space image to construct the low rank constrained coefficient matrix Z in advance.Then a linear self-representative model that's built by label coefficient matrix to exploit the label correlation is introduced to recover the possibly noisy label matrix.This introduces the matrix trace norm regularization on both feature mapping matrix and self-recovery coefficient matrix to capture the correlations between labels and also control the model complexity.In addition,using graph-based Laplacian manifold regularization as a smooth operator to incorporate the unlabeled images into the training images and can explicitly take into account the local geometric structure on both labeled and unlabeled images.To pledge the effectiveness and efficiency of the proposed model,empirical experiments were conducted on five different widely available image datasets to demonstrate the effectiveness and efficiency of the proposed framework.