基于张量罚偏最小二乘的自动图像标注

如何有效地挖掘变量与标签之间的相互关系和处理高维数据是自动图像标注的两个具有挑战性的问题。以往的自动图像标注都是基于向量模式的学习算法,这样一方面产生高维数据,另一方面破坏了图像数据的高阶结构和内在相关性,导致信息丢失。向量模式下的罚偏最小二乘算法(penalized partial least square,PPLS)可以在获取变量和标签相关性的同时,进行维度约简。在PPLS的基础上,提出基于张量罚偏最小二乘算法(tensor-PPLS)。首先构造图像的张量数据形式,然后采用多线性主成分分析(MPCA)进行降维预处理,最后用tensor-PPLS进行图像标注。在图像标注的三个标准数据集上,提出的算法标注结果明显优于传统的基于向量模式的学习算法。

A penalized partial least square algorithm based on tensor for automatic image annotation

How to effectively exploit the correlations of variables and labels, and tackle the high-dimensional problems of data are two major challenging issues for automatic image annotation. The previous methods are usually based on vector, which extract the underlying characteristics of the image. However, this may result in the following problems: (1) generate the high dimensional features; (2) break the inherent higher-order structure and correlation of the image, and lead to the loss of information. Penalized partial least square(PPLS) can capture the correlations and perform dimension reduction at the same time. Based on PPLS, we propose tensor penalized partial least square(tensor-PPLS) to solve automatic image annotation. Firstly, each image is expressed as a tensor. Then multilinear principle component analysis(MPCA)is used for dimension reduction. Finally, we use tensor-PPLS to train model and annotate the test image. The experimental results conducted on three benchmark data sets show that the proposed method is promising and superior to the state-of-the-art automatic image annotation methods.