Research On Image Retrieval Based On Multi - Feature Weighted Fusion

The rapid development of network technology and multimedia technologypromote the fast exchange of information as well as change the way people access toinformation, its application fields are also growing. Due to the continuous renewal ofinformation acquisition way, the uncertainty of information itself and othercharacteristics, how to extract images which conforms to the requirements from thehuge quantity and changing updated database has become a spot in the study, all kindsof image retrieval methods emerge in endlessly. In the early days, the more commonimage retrieval method was based on image annotation which from the text. Howeverbecause of the human subjectivity, combined with the hard implementation ofartificial annotation in a limited time, this algorithm can not satisfy the needs ofpeople. The rapid development of the content-based image retrieval methodfundamentally solves the problem in image retrieval. This paper’s main research workis that puts forward a kind of image retrieval method based on multiple featuresweighting fusion. This method is based on the continuous improvement of the imagesegmentation, feature extraction, feature fusion technology. Moreover, the main workof this paper is as follows:1. The image target recognition application based on the combination ofwatershed algorithm with level set algorithmThis method using the higher advantages of watershed algorithm in imagesegmentation precision, combined with the advantages of level set algorithm, andputting forward a target recognition method based on region merging. First, watershedalgorithm is used into original images and segment these images into multiply smalland continuous areas, then coupled with level set algorithm to merge these areas.Finally complete the target image recognition. This method is not only avoids theover-segmentation phenomenon on the vision, and greatly improves the precision oftarget recognition.2. Quantize color feature vector, improve LBP algorithmIn order to get more accurate image retrieval results, in terms of color featureextraction, in this paper, a quantitative method of segmented color feature vector ispointed out. This method realizes the change of characteristic vector from highdimension to multi-dimensions, thereby reduces the time complexity of characteristicscalculation. Moreover, based on the rotation invariance of textural feature, this paperalso makes the improvement to the traditional LBP algorithm, and joins scaleinvariant and rotation invariant into the LBP coding which better describes the image texture features.3. Implement multiple features weighting fusion method for image retrievalThe methods of target recognition and eigenvector extraction which pointed outin this paper applied to image retrieval, and then the color features and weightedfusion of liberal arts could be realized. Moreover, the traditional way of weightingmethod which based on fixed weights is also improved. Therefore the new weightingscheme based on weight matrix could be put forward. The distance betweendimensions based on the eigenvector confirms weight of size and form into weightmatrix, and then according to the size of the order of the similarity between featurevectors to complete the image retrieval work. In order to verify the effectiveness ofthis method, other related approaches also were compared and find that this wayimproves the recall ratio and precision ratio. In addition, because of the weight matrixis not for a particular data set, this method also has retrieve robustness when there is achange in data set.