多标签类重力密度和距离的图像注释方法

Image Annotation Approach Based on the Density and the Distance of Multi-label Gravitation

图像注释是一项针对网络图片的语义理解应用,由于图片的基本视觉特征与人类图像识别之间存在着巨大语义鸿沟,为此本文提出一种基于多标签类重力密度和距离的图像注释方法.现有的图像注释数据集多由多个标签组成,然而现有多标签学习算法,有着待定参数过多、精度不高的问题,因而本文通过引入类重力模型的方法,将图片之间的相似度转化成物理学领域里的重力,借助类重力密度与距离公式,选取图片训练集中若干相似度较高的图片,再对其有关标签进行处理,从而找寻图片与图片之间所隐藏的潜在关系,实现对相应标签的注释.通过对比相关经典算法,在三组基准多标签数据集和两组图像数据集上良好的实验结果表明该算法可适用于不同类型的多标签数据集,是一个可筛选出类别相近例子的多标签图形注释算法.

Image annotation, which is usually structured as a multi-label learning problem, is one of important tools used to enhance the semantic understanding of web images. Many multimedia applications can greatly benefit from image annotation. However, there exists the huge semantic gap between the available image features and the keywords that people might use to annotate. Hence ,this article proposes a new image annotation approach based on the density and the distance of multi-label gravitation （DDG）. Nowadays, the existing image annotation datasets usually consist of various multi-labels, but most of them have the weaknesses of complex parameters and unsatisfactory classification results. By simulating the law of gravitation in the physical world,the proposed method transforms the gravitation of objects into the similarity of images to overcome those difficulties. Owing to the density and distance of multi-label gravitation （ DD ） formula, DDG is able to select several most similar images from the training dataset. On this basis, DDG adds the aggregation concept to expand each image to a cluster,and searches the potential correlation among images by dealing with the corresponding label space of the cluster. According to the weight mechanism on the cluster,DDG is able to give the final multi-label results to predicted images. The multi-label gravitation model ensures the rationality of image annotation. Moreover, superior experimental results on two images datasets as well as three benchmark multi-label datasets will demonstrate the effectiveness of DDG, which is brief and easy to implement, in comparison with the state-of-the-art method.