Font Size: a A A

Research OnImage Discriminative Representation For Image Retrieval

Posted on:2014-03-28Degree:MasterType:Thesis
Country:ChinaCandidate:S Y SongFull Text:PDF
GTID:2268330428462254Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
With the development of the Internet, the rise of social media and the popularization of image acquisition devices, a large number of images have emerged on the Internet.The explosive growth of images bring great challenge for image retrieval. In image retrieval, the Bag-of-words model is usually first used for image description,then the returned images are proposed by RANSAC(RANdom SAmple Consensus) for geometry verification to RE-RANK, that usually give a better result. This image retrieval framework has3shortcomings:1) Spatial information of images is ignored completely, thus the description is less discriminative;2) For large scale image retrieval, a great amount of visual vocabularies are needed for image description. Therefore the computation cost is high if the Bag-of-words model is used directly.3) RANSAC-based geometry validation or image verification is time-consuming, due to the computation complexity. Item2and item3lead to inefficient image retrieval.To overcome the shortcomings mentioned above, this paper is focused on improve the performance of image retrieval, by using spatial information. This paper also further the study on natural scene Chinese character recognition using image retrieval method. Works in this paper is mainly covers2aspects:1) This paper designs a retrieval framework with identifying spatial information:on the first layer, spatial min-Hash algorithm is proposed using coarse-grained geometry information. Hash representation is zero-order approximation to Bag-of-words model. It randomly selects and compares a subset of visual words represented by Bag-of-words. The computation is speed up while scarifying some discrimination. In order to increase the recognition performance of Hash representation, this paper will first use spatial pyramid to present images and then process each part of the image by the min-Hash method to improve retrieval performance. On the second image verification layer, fine-grained geometry information, presented by detailed spatial pyramid, is used for verification between images. MSER regions and spatial relationship between angles and points are used for matching and verification. The verification avoids exact matching between all points of images. It reduces computation amount and accelerates verification by layered verifications of features. 2) This paper adopts image retrieval techniques in Chinese character recognition in natural scenes to address problems of font inconsistency, dataset imbalance, variety of Chinese character categories and shortage of samples within a category etc. Iterative quantization is used in Chinese character recognition and the recognized result will be rectified with the help of editing distance.
Keywords/Search Tags:min-Hash, spatial min-Hash, local spatial pyramid, iterative quantization
PDF Full Text Request
Related items