Research Of Context Based Image Copy Detection

Copy image refers to a new image which is obtained by cropping,scaling,rotation,changing compression rate,adding other content,and other non-affine geometric transformations on original images.Copy images have a common characteristic: some or all parts of copy images come from its original image.Copy image not only brings the copyright problem of image on the internet,but also occupies a large number of storage spaces.Image duplication storage is causing the waste of resource.How to quickly and accurately find the copy image is an important research direction in the field of image retrieval.Currently,the state of the art method for copy image retrieval is based on Bag-of-Words(BOW)model.In this kind of methods,an image is represented as bag of local features,and then local features are quantized to visual words.However,visual words have much less discriminative power than text words because of quantization loss.The loss of quantization on local feature causes a large number of mismatching local features which influence the precision of copy image retrieval.Aiming at the disambiguity of the visual words,a contextual descriptor based copy image retrieval method is proposed in which contextual descriptor consists of neighboring local features and their spatial relations.If a candidate match of local features has similar neighboring local feature,we consider the pair of local feature is a true match.The processing steps of this method are following: First,the Euclidean distance and scale of local feature are used to select the neighbors as the context of a local feature in an image.And then,the information of neighbors,such as position,dominant orientation and visual word is used to construct contextual descriptor.Finally,the similarity between images is measured by the number of true matches of local features.This method belongs to pre-verification method,and it takes less time in image retrieval.In order to improve the retrieval performance,a global context based copy image retrieval method is proposed.The processing steps of this method are following: First,candidate images are extracted,which are gotten from retrieval inverted index files.Then,the other matched local features in the candidate image are considered as context.During the features verification,they are verified by voting manner.If a considerable number of contextual local features are consistent with verified feature,the matching pair of local features is considered as true match.Finally,the total number of votes obtained from the true matched features is used to measure the similarity between the images.And after being verified,the method can do achieve better retrieve results.