Automatic image tagging by multiple feature tag relevance learning

Abstract

In this paper, we present an image tagging framework based on multiple feature tag relevance learning (MFTRL). First, in specific feature space, each training image is encoded as a sparse linear combination of other training images by ℓ 1 minimization, component images are treated as the nearest neighbors of the target image, so we can get each image's ℓ 1 nearest-neighbor by the ℓ 1 norm cost function. Then, maximum a posteriori (MAP) principle is utilized to determine the tag relevance for the testing image in specific feature space. Finally, the output of many tag relevance by diverse features can be combined in the manner of combining multi-feature tag relevance. The experiments over the well known data set demonstrate that the proposed method is beneficial and outperforms most existing image tagging algorithms. © 2012 Springer-Verlag.