The Research Of Image Search Techniquetion Based On Content

With the development of the times and the popularity of Internet technology, image with its intuitive, easy use have been widely used in all walks of life. However, the growing number of images brings great difficulties to the organization and query about images. Image retrieval, as an important means to solve this problem, has been widely studied and successfully applied to the object recognition, medical, military advisor, aerospace, remote sensing analysis. In recent years, the surge in the number of image database make the traditional image retrieval method impossible with regard to speed, accuracy, memory consumption, therefore researches on large scale contentbased image retrieval is very important.In this thesis, according to the problem in content-based image retrieval, we focus on global feature based-on and local feature based-on image retrieval method, aiming to extract distinctive and robust image descriptor so that user can search for similar picture give a query image. Our method can be applied to copyright protection, information security. We propose the block color histogram descriptors CTDesc(Color and Texture Descriptor), local descriptor ORB-DM(Oriented FAST and Rotated BRIEF) and heuristic reranking algorithm HEUSAC(Heuristic RANSAC). The main work of this thesis include:First, we propose block color histogram descriptor CTDesc for global image retrieval. Traditional histogram methods only consider the distribution of colors and texture in the image, while the location information is ignore, so it can not cope with image cropping, image blocking attacks. We divide the image into different levels for block, and extracts a color histogram of the block color histogram to establish descriptor, this method gets higher accuracy than traditional one.Second, in order to find user-interested local image, this thesis applies ORB-DM to extract image descriptor. Traditional intensity comparison based on methods consider that each bit of descriptor has equal importance, however some of them are easily changed because of the image transformations. In order to solve this problem, this paper proposes ORB-DM algorithm, which proved by experiments has better robustness and discrimination compared to the state-of-art methods.Third, this thesis proposes a heuristic reranking method HEUSAC. To improve the accuracy of image retrieval, geometry verification is a common method. However, the traditional method uses a random sample method which can lead extra computation. We propose a method that selects the sample points according to the results of matching, which only need to run once.As a conclusion, based on thorough analysis of the limitation of tradition method, this thesis study three important problems: block color texture descriptor CTDesc, local binary descriptor ORB-DM and heuristic ranking method HEUSAC. Consequently, our method provides solutions to large scale image search and has promising prospect.