Image Retrieval Combining Attention Model With Semi-supervised Learning

In recent years, content-based image retrieval (CBIR) becomes a hot research topic in multimedia. So far, some mature image retrieval systems have been proposed. However, due to their inherent problems, such as: human visual model can not be imitated perfectly and the inner connection among features of images can not be extracted deeply etc, hereby the retrieval results are always not quite satisfying.By analyzing the key problem of CBIR, this paper proposes a novel image retrieval method combining attention model with semi-supervised learning. The contribution of this paper is summarized as follows:(1) Salient region extraction: Based on the visual attention model, firstly we adopt the Itti's model to obtain the general saliency map, then OSTU (Maximum Inter-class Variance) method is used to get the segmentation map. Finally, the salient regions of images can be computed by combining the saliency map with the segmentation map.(2) Image retrieval based on salient region: once given the salient regions, a new image retrieval algorithm based on the attention model is proposed, which can achieve better retrieval results.(3) Using semi-supervised learning algorithm to improve image retrieval performance: In order to extract the inner connection among features of image, manifold-ranking is used to refine the initial retrieval result.From the experimental result, salient region of a given image can be extracted effectively, and the result can be naturally adopted in image retrieval. We can draw the conclusion that image retrieval combining attention model with semi-supervised learning can significantly improve the performance of image retrieval.