Sparse strategy of sample labelling for feedback in content based image retrieval

In content-based image retrieval, the retrieval performance heavily depends on two related aspects: the description of image content, and the distance function applied for sorting. In this paper, to bridge the gap between the low level visual feature and the higher semantic concept, a modified pseudo semantic model based on ensemble learning is first built for the effective description of the image content. In each round of feedback, the sparse strategy of sample labelling is exploited, thus those most informative samples can be provided to the user for labelling, which makes the subsequent online learning abilities improved for next round of feedback, i.e. the learned classifier can take on stronger generalities.