Research On Robust And Discriminative Copy Detection Technology

With the rapid development of network communication and multimedia technology,digital multimedia content is getting easier to be illegally replicated,modified,and distributed.Therefore,the problem of copyright piracy and infringement is getting more and more serious.How to protect the copyright of digital multimedia content has become a crucial and hot issue in the field of the protection of intellectual property rights.Content-based copy detection(abbreviated as:copy detection)directly extracts features representing the multimedia content itself,and then compares the extracted features between different multimedia content to identify illegal copies.Compared with digtal watermarking technology which is also used for copyright protection,the main advantages of copy detection technology lie in that any addtional information is not needed and feature extraction can be implemented after distribution.In addition,copy detection can be also applied to other emerging applications including redundancy elimination,news video tracking,etc.Therefore,copy detection technology has important significance in the filed of copyright protection and data management of multimedia content.An ideal copy detection system should have the basic properties of robustness,discriminability and efficiency simultaneously.In order to obtain the desirable balance between robustness and discriminability,the extraction of robust and discriminative features is the key issue of copy detection.This dissertation will study image and image sequence(also called video)copy detection technologies,and concentrate on the extraction of robust and discriminative features.The main contributions of dissertation are presented as follows.(1)The existing global features used for image copy detection are usually discriminative.However,they are sensitive to some geometric transformations,especially to common rotation.To address this problem,the global features based on rotation invariant partitions are proposed for image copy detection.Firstly,according to the intensity orders of the valid image pixels,these pixels are divided into several image groups,and thus each group corresponds to an image partition.Secondly,two global features are extracted from these partitions from image gradients.Finally,the extracted features are compared between images to implement copy detection.Due to the rotation invariance of the intensity orders,the constructed image partitions also up to 92%,which are better than that of the existing global or local feature based image copy detection methods.(4)In order to futher improve the effiency with maintaining the high effectiveness,a novel video copy detection method based on the combination of multiple features and spatio-temporal filtering is proposed.In feature extraction stage,three local features are extracted from the gradient statistical property,the texture information,and structure information of frames,respectively.Then the model of "bag of words" is employed to represent the extracted features as "combined words",and the inverted indexing structures of visual words is constructed,which facilitates the search of combined words so as to efficiently detect matches of the reference candidate videos.Finally,a spatio-temporal filtering strategy is employed to eliminate the false matches between candidate videos to futher improve the accuracy of detection result.Experimental results on the video datasets from TRECVID 2008 demonstrate that the proposed method not only has high efficency,but also achieves high accuracy for detecting the video copies generated by various common video copy attacks.In conclusion,this dissertation has proposed four copy detection methods to study how to address the problems of robustness,discriminability and efficiency in the existing copy detection methods.The proposed methods can effectively and efficiently detect the copies generated by various common copy attacks,and provide novel methodology and key techniques for copy detection.Research achievements in the dissertation will offer basis and guarantee for the copyright protection and data management of multimedia content,and be helpful to promote the healthy and orderly development of digital media industry.