Image Retrival Method Based On SIFT Feature And Distance Metric Learning

With the change of societyand the way of production and life, the number of image data becomes larger and larger on the network. Content-based image retrieval(CBIR), in recent years, as an important mean of image data processing, attracts attention widely. It is a challenge that extending CBIR to a large-scale image information database, enhancing retrieval performance, and meeting the needs of users. At the same time, it also is an opportunity.This thesisfirstly focuses onthe representation and description of images. In the study of CBIR,Scale-invariant feature transform(SIFT) is commonly used to represent images.In order to make the representation more streamlined, in the system established in this paper, image representations use algorithms based on SIFT and principal component analysis(PCA). Faced with the massive image data, we propose that establishment of parallel computing image retrieval process. Clustering of the database isthe foundation for building distributed retrieval. So we propose a clustering algorithm that is based on the spectral clustering. In retrieval, we propose that constantly measure the distance learning based on the feedbacks of users to improve the accuracy of retrieval.The main work of this thesis is as follows:1. This paper describes the principles and SIFT feature extraction process,and use an algorithm based on principal component analysis to extract the most streamlined and powerful features vector of images.2. This thesis attempts to solve the problem of massive images in images retrieval, by founding a distributed indexing. In other words, we clusterdata of an image database as the basis, and thenbuildk-Dimension tree(KDtree) on the MapReduce paradigm to process in parallel.3. To divide the dataset, this paper researchs on clustering algorithms. We analyze defects of the existing algorithms, the spectral clustering methods and density-based clustering method,and propose a clustering method that clusters data twice in one process.4. In order to achieve better search results, we use the distance metricearning algorithm for learning a distance metric from the labeled images.Combined with the feedbacksabout constantly intermediately results of users, this paper establishs an interactive retrieval model based on distance metric learning.The extracted feature vectors still has translationinvariant, scale invariant,rotationinvariant and affineinvariant.Improved clustering algorithm has a high perforamce. And it also is the basis for distributed search strategy, which solves the challenges brought about by large-scale image retrieval to a certain extent.Based on the distancemetric learning, our interactive retrieval brings an enhanced on userexperience.