Research Of Image Annotation Algorithms And Its Implementation On Hadoop Platform

Machine learning is an important and tough task in computer science research,and learning based on visual information is called computer vision.Automatic image annotation is a research hotspot in computer vision,which is mainly used to deal with the massive image management and intelligent retrieval of the Internet.The traditional image annotation is done manually,but the efficiency of manual labeling is too low to meet the requirements of the current because of the "big data" age.So computers must learn the corresponding image labeling rules to achieve more efficient automatic labeling.And the focus of automatic image annotation is both the high efficiency and accuracy.In order to improve the accuracy and efficiency of image annotation,image annotation with respected to massive images is studied in two aspects.The accuracy of the image annotation based on Positive-Negative Instances Learning is improved by adding a filter and the corresponding distance measurement.Designing a new image annotation method based on BHF-LDA(Fast LDA Based on Hadoop),which reduces system runtime greatly.The main contents of this thesis are as follows:(1)Image feature extraction methods are studied,and color feature,texture feature and SIFT(Scale-invariant feature transform)feature of the image are extracted.The N-cuts algorithm and Mean-Shift algorithm are studied,image segmentation based on N-cuts and Mean-Shift-N-cuts are designed and implemented.The experiments show that image segmentation based on Mean-Shift-N-cuts is much more efficient and practical.In addition,bag of visual word model is studied,and the evaluation standard of image annotation is introduced.(2)Image annotation based on nearest neighbor is designed and implemented.The experiments show that the choice of feature has great influence on the image annotation results,and the JEC(Joint Equal Contribution)distance is obviously better than the other features.Image annotation based on Positive-Negative Instances Learning is designed and implemented,and we have experimented many times by changing positive and negative weights.For lack of efficienct combination of Positive-Negative Instances vectors,image annotation based on Positive-Negative Instances Learning is improved by adding a filter and the corresponding distance measurement.The experiments show that the improved image annotation method improves the accuracy greatly.(3)Focusing on massive images on the Internet,a fast image annotation method is proposed.The structure and principle of Hadoop are analyzed and a Hadoop-based platform is built.LDA(Latent Dirichlet Allocation)model is studied,and the parameters of LDA are solved by Gibbs sampling.Image annotation based on LDA is designed and implemented.The experiments show that this method can achieve high precision.A new BHF-LDA based algorithm is proposed,which utilizes the Mapreduce parallel computing framework to save LDA training time.Experiments show that compared with the single-point environment,image annotation based on BHF-LDA greatly improves the efficiency of the labeling system.