Image Retrieval Using Salient Points And Keyblocks

In this paper, a novel method is proposed, which is based on salient points and key-blocks. First we extract salient points using wavelet transform, thus segment the image into equal size blocks. These blocks can be divided into two classes, ones with the salient points and the others not. Then low-level features such as color are extracted to describe block. By adjusting the weights of the two classes in image retrieval, we can get different results which depend on our interest in local or global similarity.Wavelet-based salient point has been used for image retrieval. Experiments show that the salient points are mostly found on the boundaries of the objects, therefore the local features around the salient points primarily represent the object information instead of the background. On the other hand, keyblock-based approach also provides a practical solution for content-based image retrieval analogous to text-based IR techniques, but the keyblocks are constructed on the basis of the entire image, so they mainly contain the global features of the image. In real world, because the viewers have different focus on the same image, thus the wanted retrieval results may also be different. Therefore, we hope to integrate these two methods effectively, so that we can not only improve the retrieval efficiency, but also get more degree of freedom during the retrieval.We have performed a series of experiments to confirm the effectiveness of our method. Experimental results indicate that our method not only owes the advantage of keyblock-based approach , but also the salient points. Moreover, users can get satisfactory retrieval results according to their interest by adjusting the weights.