Research On Particular Object Based Large Scale Image Retrieval System

With the development of computer technology, multimedia technology and the internet technology, more and more people come into contact with the mass of the image information. How to retrieve the required image by the user from the mass is a very important issue.Research on particular object based image retrieval technology is an important branch of the image retrieval field, which has become a hot topic in recent years. With the development of internet and the improvement of hardware, image database is becoming larger and larger. Current methods for retrievaling could not meet the demand of users in accuracy and efficiency. To solve this problem, we mainly focus the key technique related with accuracy and efficiency. The main contributions are following:1) An improved RANSAC algorithm is proposed. In order to deal with the high outlier ratio of RANSAC, it adopts the weak geometry hypothesis to filter the data. Experimental results show that the proposed method is efficient and accurate.2) A large-scale image retrieval system for particular object is proposed, which uses Principle Component Analysis to reduce the dimension of feature descriptors; uses approximate nearest neighbor method to scale traditional k-means to large vocabulary, uses the proposed RANSAC method to increase the speed of spatial verification; uses query expansion to boost recall.3) Two typical applications are presented:book recognition system on robot; mobile movie retrieval system based on posters...