Manifold Learning Dimensionality Reduction from SpectralTheory for Automatic Web Image Annotation

Automatic image annotation is one of the most challenging problem in machine vision. The goal ofthis task is to predict number of keywords automatically for generic images captured in real environment.Many methods are based on visual features in order to calculate similarities between image samples. Butthe computation cost of these approaches is very high. These methods require many training samples tobe stored in memory. To lessen this burden, a number of techniques have been developed to vastly reducethe quantity of features in a dataset—i.e. to reduce the dimensionality of data. Manifold learning is apopular approach to nonlinear dimensionality reduction. In this paper, we investigate manifold learningmethods from spectral theory for image auto-annotation task. Laplacian Eigenmaps (LEM) and Hessianlocal liner Embedding (HLLE) are used to reduce the dimension of some visual features. Extensiveexperiments and analysis on various datasets and different visual features show how these simplemanifold learning dimensionality reduction methods can be applied effectively to image annotation.

کلیدواژه ها:

Automatic image annotation ، dimensionality reduction ، manifold learning ، spectraltheory