Research On Image Retrieval Based On Comprehensive Features And Salient Points

With the large capacity storage equipment and digital equipment coming into vogueand rapid spreading of multimedia technology and network technology, image hasbecome a kind of common and important information carriers, and are growingexponentially. Image retrieval technologies are presented to retrieval the expectedimage information in the vast image database, and it has been developed through twostages: text-based image retrieval(TBIR) and content-based image retrieval(CBIR).CBIR helps people get rid of the boring and heavy annotation task of text-based imageretrieval. CBIR attracts more and more attention, and is successfully applied to medical,education, the digital library, the military, industrial and commercial, etc.Many problems still need to be further researched in CBIR. There is no kind offeature that is suitable for representing the content of all of the images. It’s gettingmore and more important to meet the increasing real-time performance demand ofimage retrieval in large-scale image database. Constructing the mapping fromlow-level visual feature to high-level semantic feature is still an open problem. Thisthesis research on CBIR, mainly focuses on image retrieval based on comprehensivefeatures and image retrieval based on salient points. The main content and researchresults are as follows:①The research background, research significance, and development of imageretrieval are introduced firstly. Extraction algorithm of the low level feature, similaritymeasurement and standards of image retrieval performance are described in detailedafter the research status and hot topics of CBIR been discussed.②An image retrieval method based on comprehensive features is proposed.Feature extraction is the key step of image retrieval. This thesis researches on theclassic feature extraction algorithm, and mainly focuses on texture feature—blockdifference of inverse probabilities(BDIP for short), and color feature—linear blockalgorithm(LBA for short). Taken as an image feature, BDIP moments is used toconstruct a feature vector whose dimension is very high. Furthermore, using the meanvalue of BDIP to classify the blocks may result in grouping the blocks that belong tothe same object into different classes. To solve the problems above, this thesis proposesa new texture feature—dominant block difference of inverse probabilities (DBDIP forshort). Firstly, it extracts the BDIP image of the original image, then obtains the texture feature by calculating the dominant colors of the BDIP image. Image retrieval iscarried out by combining LBA and DBDIP. The experiment result shows that DBDIPrepresent the image content very well, and the proposed method gets higher precisionand recall rates than the classical methods.③Image retrieval based on salient points is proposed by using an auto extractionalgorithm of salient points. The salient points extracted by traditional algorithms aredispersing, and the time complexities of the traditional algorithms are high. The autoextraction algorithm of salient points proposed by this thesis can solve these problemseffectively. The proposed method use BDIP model to extract the valleys and edges,then salient points are extracted by iterative threshold selection method. Theexperiment result shows that: the salient points can be extracted efficiently and thehigher rates of precision and recall can be obtained when the difference between theforeground and the background is obvious; on the other hand, the salient points can beextracted with low efficiency and the rates of precision and recall will be reduced if thedifference is not obvious.