Design And Implement Of Image Retrieval System For Internet Sensitive Content

With the development of the Internet, the Internet brings huge convenience to ourlives, and has been becoming an important way to communicate with each other.However, the Internet also brings a huge challege to us. On the Internet there is lots ofharmful information related with eroticism, violence, reactionary, heresy and so on.This type of information not only harms the physical and psychological health ofchildren, but also seriously threatens the national security and the social stability. It isnecessary to build the public opinion monitor and warning strategy. In the recent years,mass disturbances spring up. Ihe Internet has been becoming one of important reasonsto expand the situation of the mass disturbance. To search and monitor the sensitiveinformation on the Internent is an efficient way to early warn and prevent the spread ofthe negative influence.An Image retrieval system for Internet mass disturbance sensitive content isdesigned and implemented in this thesis. By using this system, we could find the similarimages and their webpages related with mass disturbances on the Internent.Since theamount of images is very huge and the number of feature dimensions is very high, theimage retrieval system would respond slowly. The feature extraction methods have beenimproved. The number of feature dimensions was reduced and optimized by usingdimension reduction techniques. The speed of system responding was improved. ThisThesis also described the designment and implement of the image retrieval system forthe sensitive content on the Internent.Firstly, the thesis described the requirement analysis and the general designment ofthe proposed system, and introduced the software and hardware situation of the systemand its development platform. Based on the analysis of system performance and systemfunctions, we proposed the general rule of the system development, and summarized themain functions of the system. In particular, the thesis described the workflow andimplement idea of the topic network spider and the designment of the database of thesystem.Secondly, we researched the algorithms of image feature extraction, and proposeda feature extraction method based on graph embedding framework of dimensionreduction. The number of feature dimensions was reduced and the subset of imagefeatures was optimized by using the dimension reduction technique, then the betterretrieval result would be obtained. Comparing with the existing main mehods, thevalidity of the method proposed in this thesis was verified on the international testingimage set. Meanwhile, this thesis introduced the technique of user feedback. The userscan grade the resulting images, and the similarities of images were resorted by using thefeedback of the users. This procedure could be repeated until the satisfied result was obtained. The retrieval result was expected to be opitimized by this strategy. And thestrategy was expected to capture the true intention of the users.Thirdly, the thesis described the detailed designment and implement of the system.Based on the object oriented programming method, we introduced the designment ofmain classes of each module。 We also provided the main codes for image color featureextraction and texture feature extraction. With the practical images of mass disturbances,we showed the user interfaces and using flows of the system, provided the retrievalresult of the system.Fourthly, the thesis researched the system optimization design of GPU based onhardware acceleration. We introduced the CUDA system architecture, and detailed ofbuilding CUDA+VS2008environment. Considering the system parallelism analysis, wetransplanted calculation module of HSV eigenvalue to the CUDA. Through theexperimental test, it verified the feasibility of this method.Finally, the thesis summarized the main implemented work and disadvantages ofthe proposed system, and proposed possible directions for further research.