Font Size: a A A

Research And Implementation On Multi-label Learning-based Image Annotation

Posted on:2018-09-04Degree:MasterType:Thesis
Country:ChinaCandidate:J W LiFull Text:PDF
GTID:2348330512471727Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
In recent years,under the background of computer technology,digital media and the rapid development of multimedia information,peoples' life is greatly involved in mobile phone,digital camera and other high-tech products,which leads to continuously explosive growth of massive image information.Therefore,how to effectively retrieve images from massive images has become a hot research issue in the fields of computer vision and image processing.Automatic image annotation algorithms assign sematic related tags on a given image automatically,and it has become an important way on image classification and image retrival.Most studies treat image annotation as a typical multi-label classification problem which brings progress to automatic image annotation.However,they still exist some drawbacks in the face of massive image data.First,the number of training samples is limited in reality.Second,most tags on labled images are noisy.The last one is how to make the predicted keywords reflect image feature accurately.The three problems has become the most important limiting factors on accuracy of image annotation.In order to improve the accuracy of image annotation,this paper makes a classification and deep analysis on existing image annotation algorithms and then solves this problem from two sides on the basis of multi-label learning,which are semi-supervised learning method combined with low rank constraint and tag ranking.This paper proposes two image annotation algorithms and proves the high efficiency on ESPGame,IAPRTC-12,NUS-WIDE and other databases.The main research results of this paper are as follows:(1)Image Annotation Algorithm Based on Structured Low-rank and Semi-supervised Frame.This algorithm concatenates the prediction models for different tags into a matrix,and introduces the matrix trace norm to capture the correlations among different labels and control the model complexity.In addition,by using graph Laplacian regularization,the proposed approach can explicitly take into account the local geometric structure on both labeled and unlabeled images.Moreover,considering the tags of labeled images tend to be missing or noisy,we introduce a supplementary ideal label matrix to automatically fill in the missing tags as well as correct noisy tags for given training images.This paper gives the solution to the algorithm frame,and conducts image annotation experiments on a variety of databases which demonstrates the effectiveness of the proposed approach.(2)Image Annotation Algorithm Based on Tag Ranking and Matrix Recovery.This approach ranks tags in the descending order of their relevance to the given image instead of having to make a binary decision for each tag which solves the problems of limited training samples and incomplete labeled image tag collection.Moreover,this approach casts tag ranking into a matrix recovery problem and takes matrix low-rank into account which restricts the correlations among different labels.So that a reliable prediction model can be learned for tag ranking even when the tag space is large and the number of training images is limited.This paper employs Accelerated Gradient Algorithm(AGA),and demonstrates the effectiveness on databases Corel5K,ESPGame,IAPRTC-12 and other datasets.
Keywords/Search Tags:Automatic image annotation, Multi-Label Learning, Graph laplacian matrix, Semi-supervised learning, Low-rank, Trace norm
PDF Full Text Request
Related items