Research On Large-Scale Content-Based Image Retrieval

Actually the image contains richer and more effective information than text, it takes a very important role in daily life. In modern life, large number of digital image data is explosive growing. In order to retrieve the image accurately, it is necessary to use the real content of images, which means using their characteristics such as color, shape, texture and others. That is the definition of CBIR (Content-based Image Retrieval). Because of its practical value, researchers around the world take more widespread attention to CBIR.In general, CBIR is essentially to solve two core issues: feature representation and extraction; similarity measure. This paper will be focused on these two core issues and how to extend CBIR to large-scale data, and do research on how to describe the content of image, how to extract features accurately and automatically; how to get an accurate similarity result; and how to work quickly on large-scale and high-dimensional feature data.(1) First based on the current image feature extraction technique, the article does research on how to improve the segmentation and de-noising effection to make it more suitable for network image, then extract meaningful regions and key regional characteristics based on fragmentation and de-noising technology. Then the article proposed a novel method named stacked-triangular geometry histogram that can representation the feature of image more effective and solve local-deformation problem better.(2) Then the article proposed and implemented a new vote-based index and similarity measure method. This new algorithm unified two processes more efficient and achieved a balance between speed and efficiency. Later, it also implemented a framework based on this idea of a high-dimensional feature indexing and retrieval which can be used for other similar situations.(3) At last, based on the above principle and idea. We design and develop a general large-scale content-based image retrieval engine and platform named Magic with a strong expansibility. The platform facilitates the researchers to do the experimental study of image retrieval and improve the work efficiency. It promotes the academic exchanges, and this work has great practical value.