Key Technology Of Instance Search In Image And Video

With the rapid development of computer and network technology, a large number of images and videos are produced in our daily life. In order to mine out the information needed, these media materials need to be processed quickly and accurately by scientific technology, instance search has a very practical value in this respect. In this paper, we would carry out a detailed research on the key technology of instance search in image and video, the main work is as follows.Aiming at the shortcoming of the traditional BoW model, which ignores the spatial information of the feature points, as well as the existing RANSAC-based methods with the disadvantage of time consuming, we design and implement a much faster approach named direct spatial matching making full use of the location and scale information of feature point. Compared to traditional BoW-based methods, which direct use the descriptor information of SIFT feature point, we got more improvement in the result. We also use a K-NN reranking scheme to further improve the accuracy.We construct a region proposal based depth features for instance retrieval. We explore the influence of the network types, convolution layers, encode approach and the size of input image on retrieval results.In addition, we put the region proposal idea of object detection task in to instance search. We first generate object regions on image, then extract CNN features of the corresponding region. Constructing region proposal based instance retrieval frame work, the method has achieved good results in the standard image retrieval data set.This paper propose an adaptive feature fusion method, relate the effectiveness of the feature to the shape of the score curve, processing the late fusion in score level. when test on the TRECVID INS 2015 data set,the results show that the fusion method greatly improves the retrieval accuracy. We ranked third in the fourteen teams, which verified the effectiveness of the algorithm.