Research On Content-based Binary Image Retrieval Technology

With the popularization and development of Internet, the number of image data grows dramatically fast, and how to retrieve image efficiently and quickly becomes an important issue in the field of image's application. Due to the speciality of image data, the traditional keywords based retrieval could not satisfy the need of managing and retrievaling image data. In order to manage and retrieve large amounts of images, the content-based image retrieval (CBIR) has emerged to be one of the hot research areas in image domain. It integrates several technologies in the domain of image processing, image recognizing and image database, and could provide more efficient retrieval methods.Considering the research purpose is to offer academic and technology support to icon retrieval in text images, after discussing the key techniques of image retrieval, the paper research content-based binary image retrieval. The main works are as followed:First, feature extracting techniques based on color and texture and similarity measure technology are reviewed. Then, an image retrieval system combining color and texture feature are constructed. The main color describes the whole information of image, and the edge histogram describes local information. The retrieval results obtained from combined features are superior to that obtained from single feature.Second, the contour-based and region-based shape retrieval are reviewed, mainly including Fourier, Invariant Moment, Zernike Moment, Pseudo-Zernike Moment and Legendre Moment. These five shape-based image retrieval methods are realized, and experiments are taken to compare their performance.Third, a region-based shape representation is proposed, which is call "density distribution feature (DDF)". It expresses the dimensional distribution information of the object's pixels within binary image. After shape center orientation and region partition, two M dimensional feature vectors are got. The first feature vector represents the relatively density of object pixels within each sub-image. And the second feature vector represents the difference of relatively density in the direction of radial coordinates. When matching the similarity, the paper first used the Gaussian model to normalize the distances gained from the two feature vectors, then integrated the two distances to calculate total similarity measure. The experiments results showed that this shape feature can depict the image well and is invariant to translation, scale and rotation. The paper also evaluates the effectiveness of the proposed descriptor with respect to Moment Invariants.Four, the improving methods of the CBIR systems are discussed, K-means clustering and genetic algorithm clustering are realized and applied in clustering database. K-means clustering is proposed to filter the retrieval result images, and experiments show that it enhances the retrieval accuracy.Five, relevant feedback technology is researched. And one of the relevant feedback methods is realized. It let user adjust the weights of features in term of relevant feedback so that the retrieval accuracy is improved and satisfies the user.