GMM-VLAD Based Image Retrieval

Image retrieval is the current hot research topic in the field of computer vision, it-s development has experienced text-based image retrieval and content-based image re-trieval.The latter includes image retrieval based on global features、image retrieval based on local characteristics,and now based on global express feature (essentially generated by local characteristics of polymerization).The bag of words of traditional (BOW) is a very popular and widely used global express characteristic at present. A lot of progress acquired in image retrieval is based on it. But the visual dictionary needed to construct uses large clustering number,while it brings a lot of computation cost. And due to the characteristic only contains the number of information,leading to contain less informa-tion and the ablity of the characteristic is weak.The characteristic later proposed VLAD also is a outstanding global express feature and it has a very good effect, but it also exists defect to some extent.First of all, VLAD take a hard assignment strategy when distribute the feature points,so this makes the final formation of the vector can’t express the image of the nature more comprehensive.Second, the VLAD has a high dimension,which leads to large amount of calculation when retrieving.The first major job completed of this paper is to put forward a new global express characteristic of GMM-VLAD (Gaussian Mixture Model-Vector of Locally Aggregated Decriptors). GMM-VLAD is a characteristic which proposed through a series of im-provement in the traditional VLAD with probability information.GMM-VLAD can view as the expansion of the BOW and VLAD. BOW only has the quantity information of the local characteristics of image after cluster. However,GMM-VLAD not only con-siders the local features’s distance to the center of the cluster, but also with probability information to join,it contains more abundant information.The idea is:firstly, it extract-s the image’s local characteristics to form a feature library; secondly,it fits the feature library with GMM (Guassian Mixture Model),and we get k clusters and probability density function,which we can use to calculate the probability for each characteristic to the sketch of each cluster; thirdly, in the last step we calculate the probability that we can selected from big to small order before m class with the highest probability.The weight (related to probability) of each local characteristics can be assigned to the m centers;in the end, to each image, all of the local descriptors are calculated according to the above distribution with its recent m centers,then we can get the GMM-VLAD characteristic-s. It can well solve the above problems such as the calculate cost caused by BOW and VLAD.The second job of this paper completed is to put forward a new kind of preliminary clustering retrieval algorithm,which we use pre clustering in retrieval phase that we do a preprocessing work first. We make a second cluster on the global express characteristics library in retrieval phase. To compare the query vector with clustering center instead of compared with the characteristics of each.This method can not reduce the retrieval accuracy and greatly improve the retrieval efficiency.Considering the GMM-VLAD dimension is very high, in this paper we using the principal component analysis (PCA) for dimension reduction.The experimental results show that GMM-VLAD can improve precision and recall of the retrieved results compared with traditional BOW and VLAD features, and at the same time guarantee not to increase the storage space and computation time. Theoretical analysis and experimental result show that in generally cases, the preliminary cluster-ing retrieval algorithm needs less computing time compared with the traditional ergodic search,and real-time retrieval time will decrease and it makes the user a better retrieval experience.