A Minimum Euclidian Distance Image Retrieval Relevance Feedback Algorithm Based On The Bayesian Classifier And Implementation On DSP

With the rapid development and popularity of the multimedia database, theresearch on the multimedia and related techniques becomes the hotspot of theinformation field. Especially the Content Based Image Retrieval (CBIR) methodbecomes a new forward topic. The CBIR system extracts vision characteristics suchas color, shape, texture etc. from image, and they are used as the indexing basis forimage retrieval. But in the practical CBIR system, there are significant differencesbetween semantic query used by customers and the vision characteristic descriptionof image, and it is hard to obtain satisfactory mapping relations by existinginformation processing and data analysis techniques, so users often can not meettheir demands only through the retrieval by vision features.Up to 90's in 20 centuries, the concern of vision features shrank in the researchof image retrieval. Image retrieval problem developed to the level of semanticbased image retrieval. Understanding of the users' needs for personalized contentand abstract of the image become the hotspot of image retrieval;we areapproaching the semantic from the characteristics derived from both image visionand user, trying to get the right semantic feature. From the vision feature to thesemantic feature, the gap between the two is generally considered to be "semanticgap".By studying the retrieval habits of human being, we can find that the exampleimage and the target image belong to a specifically semantic category. At present,our main research goal is to find an available way which can judge whether animage belongs to a particular category. To solve this problem, classifier wasintroduced to the image retrieval, but simplistic classifier can not simulate humanthoughts, we also can not simulate human's classification acts effectively in theprocess of image judgment. During the retrieval users are hard to express theirrequirements clearly, but when the retrieval system shows certain results, even ifinaccurate, users can easily identify with some of those they are interested in. If theuser can observe the results to the retrieval system and should not describe theirown specific needs, the retrieval system can make full use of this information toenhance the efficiency of retrieval, and shielding many details of queryconstruction. In this case users should not know the composition of the informationdatabase and content of query environment, but can find useful information, whichled to the application of relevant feedback in image retrieval. Relevant feedbackhelps us to further improve the classification of the image. In fact, the handling offeedback image is a statistical classification process. In this processing, theintervention of user (for the selection of positive and negative feedback), actuallythe feedback image processing is the mathematical modeling of user's thinkingmodel during the image judgments. Even now we can not extract image's semanticcharacteristics effectively, but by the relevant feedback based on classification wecan distinguish semantic categories.This thesis proposes a minimum Euclidian distance based image relevancealgorithm based on the Bayesian classifier to distinguish the semantic categories,and realize image retrieval task based on semantic category. The basic idea ofBayesian based reasoning is according to user's feedback adjustment on theprobability distribution of all images in the database. The retrieval process can beequal to reducing the probability of retrieval error to the least. We assume that theuser needs images and example image form a common semantic category. Then wetake this example image as a "typical model" of this category, features of thisimage is parameters of the retrieval function which is the measurement function ofsimilarity between two images. In fact we are seeking the probability of the imagebelongs to this semantic category, and this function is a Mahalanobis distance. Wedid this in experiment and found that Euclidian distance is more effective thanMahalanobis distance, so we use Euclidian distance measuring the distancebetween the two images.On the handling of feedback images, this thesis proposes a new method basedon minimum Euclidian distance. We are not doing the statistical average offeedback images and example image to update the retrieval parameters, but againstevery piece of queried image we set a relatively most suitable encouragementcenter and punishment center which are based on minimum Euclidian distance. Sowe can get the encouragement coefficient and punishment coefficient and the finalretrieval distance. Experiment results show that the retrieval algorithms improvedthe retrieval efficiency than similar algorithms.The development of network promotes the video and image applications.Excellent image quality and good real-time processing needs the increase of thecomplexity and operation of image processing algorithm, so are the higherdemands for image retrieval system. An embedded system based on DSP hasflexible programmability and higher performance and lower cost than genericprocessors thus gained extensive application. We realized this algorithm inTI'TMS320 C6711 DSK.In this thesis we propose a new approach of relevant feedback method basedon Bayesian classifier, which deserves certain referential value and practicalsignificance in promoting the development of retrieval techniques of imagedatabase.