Image Retrieval Based On Dual-space Pyramid

With the rapid development of information, The quantity of image is fast growing at an alarming rate. How to retrieve the wanted image in the large amounts of image is an important technology. The technology of image retrieve has changed from the initial text-based to content-based in recent years. The content-based image retrieve is effected largely by how the image is represented and how to measure the similarity of images. So image feature extraction、feature description and the similarity measure are a research hotspot in recent years.Now, Most of the image retrieval systems is retrieve the target image with the image database one by one. However, this method reduce the speed and efficiency of retrieval. And the existed methods for image retrieval usually use one or two image features to measure the similarity. So the result is not good.For the existing problems of image retrieval, this paper puts forward a retrieval technology based on double-pyramid-based image classification.We combine the local features(SIFT,SSIM) with global features(EDG,HSV), then put the set of features into double-pyramid (feature pyramid and image spatial pyramid)for multi-resolution decomposition. In this way, we not only consider the characteristics of the spatial relationship of different features, but also take into account the the position relationship of features. The relationship between the set of features can be accurately reflected. We compute the difference in all level in feature space of the same image area. We distribute a weight to every match result of different regions, and compute the sum of the final image matching function. Dual-space pyramid matching function is satisfied with the Mercer theorem, the matching function can be called kernel function. In this paper we embedded the dual-space pyramid kernel-function into SVM for image classification and this method has a good ability. Above all, we can retrieve the image based on the classification. For each feature, it has different measurement of similarity. So for the multi-feature,the amount of calculation will increase. This article will regard dual-space pyramid kernel-function as a function of the similarity measure, there is a one-time measure for the set of features.