| With the rapid development of multimedia and computer network technology, the publication of information has been changed gradually from single text mode to multimedia based information, such as graphics, image, animation, video and so on. At the same time, the rapid growth in the volume of information also has made Internet into a truly massive database. Human beings rely on their own organs to get the information from the world and to understand the world. According to statistics, of all the information obtained by human beings, acoustic information accounts for 20%, visual information accounts for 60%, all other information obtained from taste, tact and smell all together less than 20%. It can be seen that as the media and mean of obtaining and transferring information, image information is very important. With the wider application of image information with each passing day, people's need of retrieving image data also increases. Thus, effective management of growing image database has become an urgent problem to be solved. Flexible, efficient and accurate image retrieval strategy is a key technology to solve this problem."Information Retrieval"has already appeared half a century before, it is used to describe the process of changing the requirement of certain information into a referential aggregate, and based on the referential aggregate users may search and retrieve relevant information very quickly. Traditional image retrieval is based on text matching. The vision technology used on computer is not mature, and can not realize the automatic extraction of keywords for describing the image. Retrieval tools are very important to the success of finding the information one needed from the huge information database. Text-based image retrieval technology has been generated since the late 1970s. It first makes manual annotation to the image, then uses text retrieval technology to process keyword search. However, there are serious problems in the text-based image retrieval mode: the workload of fully manual annotation is often too heavy for people to bear; besides, the text annotation is inevitable subjective and inaccurate.To overcome these problems, in the early 1990s, the content based image retrieval technology (CBIR) was proposed. Content based image retrieval technology obtains image content in virtual of the process of processing, analyzing and understanding image from bottom level to top level, and then query by image content. Computer with this technology will automatically extract the visual features of the image content, such as color, texture, shape , and the location of objects and their correlations, etc., then it will search the images in database to match the sample image according the features, and retrieve similar images.At present, features used for the retrieval of image include color, texture, shape, and so on. However, it is not enough to use a single feature to retrieve the desired image. Because the image features automatically extracted by the computer is much different from the man's understanding of semantic meaning, the retrieval result achieved by traditional image query system which is based on the measurement of feature extraction and similarity distance is often dissatisfactory. The main cause for this is that the computer vision technology is still not perfect. The ender user of image retrieval system is human. Therefore, it is very important to capture people's cognition of the image content from a psychological point of view. In order to embed the user model into image retrieval systems, in recent years, relevance feedback mechanism has been introduced into the content based image retrieval technology,and has greatly enhanced the effectiveness of retrieval. Based on the above problems, this paper makes an in-depth study of image retrieval technology, especially to discuss the features of color, texture, and so on.The main work of this thesis is summarized as follows:First, it gives a comprehensive analysis of image retrieval technology based on the property of color, introducing the commonly used color space as well as interconversion method between various spaces, choosing HSV color space which conforms to visual feelings as research foundation. Besides, it carries out research on quantification extraction and match of the image's color features. This paper also makes a study to the image retrieval based on the main color, and improves the image retrieval methods aiming at the main color. It uses experimental methods to further improve the superiority of improving methods through selecting the means suitable for measuring similarity.Second, it introduces image retrieval technology based on the property of texture, and gives a detailed introduction to the research of texture feature corresponding to the human visual feelings on which Tamura et al. have made a series of research. Furthermore, it makes a deep research on roughness which is the most fundamental and important texture feature, proposing the idea of adopting linear quantification when selects the neighborhood size based on Rosenfeld's algorithm of texture roughness. It selects the typical texture image in authoritative Brodatz's texture image library to build texture image library and carries on experiment, which proves that improved texture roughness have a stronger resolving power of texture and that the result of image retrieval based on improving texture roughness is superior to that based on the original texture roughness.Third, it introduces the algorithm of relevance feedback on image retrieval technology in details, and proposes the improved means of image retrieval aiming towards the main color and the texture roughness, which makes the process more efficient and the result more consistent with human visual experience. It makes a key analysis on the algorithm of relevance feedback based on modifying the feature weighting, and based on which it proposed a new algorithm. Different from former algorithms, the new algorithm increases the training sample number with un-labeled image and counter-example image, solving the problems of few training sample and raising the feedback efficiency.Fourth, it takes Windows 2000 as the development platform, Visual C++6.0 as development tools of experimental system to build the image retrieval system, for the realization of basic image processing functions. It extracts image color, texture and other eigenvector, makes a matching measurement of comparability and carries on the related image retrieval in accordance with user's needs. Besides, it adds the relevant function of feedback, enhances the interactive function with users with its friendly interface and convenient operation. Practice shows that the present study fully meets the needs of image processing and the requirements of the image retrieval. To sum up, this paper completes the research on image retrieval based on image color and textural property, and add reasonable measuring methods of comparability and the design of relevance feedback algorithm, thus it can complete the task of image retrieval smoothly. The image retrieval system established can make gray-scale transformation, histogram mapping and other image processing work of the image according to user's needs, so can realize the image retrieval methods involved in this research. By adding the relevance feedback, it makes results more in line with user's needs and the human eyes'visual experience and avoids the inconsistence between underlying characteristics and the high-level semantic meaning. In the future study, we should strength the study on extraction targeting towards the feature of the image shape and further improve the algorithm of relevance feedback. We also should try our best to realize the congruity of underlying characteristics and the high-level semantic meaning, enrich system interface, enhance visibility, and strive to make the research and development of content based image retrieval system perfect. |
PDF Full Text Request