ABSTRACT

Sparse coding approximates the data sample as a sparse linear combination of several fundamental code words and then uses the sparse codes as new presentations. We study learning discriminative sparse codes using sparse coding in a semi-supervised manner with only a few labelled training samples. A novel semantic sparse recoding method is being developed to provide more descriptive and robust visual content representations for image annotation. Although it has been reported that the visual bag-of-words (BOW) representation achieves promising results in image annotation, its visual codebook is totally learned from low-level visual data using quantization approaches, leaving the so-called semantic gap unbridgeable. To address this difficult issue, we augment the original visual BOW representation by combining annotations from training photos with predicted annotations from test images. We learn the variable class labels for all the samples by exploiting the manifold structure spanned by the data set of labelled and unlabeled samples and the limitations imposed by the labels on the labelled samples. Additionally, to enhance the discriminatory ability of the learnt sparse codes, we assume that class labels may be predicted directly from the sparse codes using a linear classifier.

Keywords : Semantic chasm, sparse codes, annotation, and bag-of-words