Research On Context-sensitive Image Retrieval

With the rapid development of multimedia technology, image resources are increasingly moving onto people’s work and life. How quickly and accurately searching for their requisite information from a huge image database is one of the biggest challenges in the computer vision. To facilitate the recognition and search of the image, content-based image retrieval comes into being. The main aim of our paper is to research context-sensitive image retrieval, which be-longs to the content-based image retrieval.This paper begins with an introduction to the basic knowledge of content-based image retrieval technology. On the aspect of image feature extraction, we introduce the principle and common methods of color, shape, and texture features extraction. On the aspect of image matching method, we introduce some common similarity metric formulas, such as Euclidean distance, Mahalanobis distance, weighted Euclidean distance and so on.After understanding and mastering the basic knowledge of content-based image retrieval technology, we propose a novel framework for context-sensitive shape retrieval. This framework consists of the off-line and the on-line stages:(1) In the off-line stage, we cluster the image database into groups, and build cluster indexes for it. The paper uses a context-based clustering method. It firstly identifies a set of initial labels based on the degrees of the vertexs, and iterates an agglomerative clustering method to get a final result. In order to reduce the pretreatment time of the image database, our approach takes out part of the elements to cluster into groups, and then divides the remaining part into the groups, which is already clustered. It will reduce at least50%time consumption comparing with the original method.(2) In the on-line stage, we propose a novel framework for context-sensitive shape retrieval. Firstly, segmenting the shape feature using the articulation-invariant theory, and we get a pairwise distance metric; Comparing the query shape with the indexes of the database, we get a more similar database; Diffusing the similarity between shapes through a graph which is built with the query shape and the similar shape database. We choose three methods to get diffusion sets. They are K-NN (K-Nearest Neiborghood) method, Anti K-NN method and Consensus K-NN method. Then we get the final result.The experimental results show that the performance of the context-based image retrieval is improved when it’s incorporated with the clustering technology. We achieve a99.92%bullseye score on the popular MPEG7CE Shape-1Part B shape data set with reducing at least60%time consumption than other context-based methods.