Research On Object-Based Image Retrieval

With the rapid development of Internet and multimedia technology, multimedia information such as images has exploded. The traditional text-based information retrieval technology has been unable to meet the needs of image information retrieval, and content-based image retrieval has become an important research area, while object-based image retrieval has been the hotspot in this area.In the field of object-based image retrieval, the current mainstream approach was visual words (VW) based methods. But these methods neglected the spatial relationship among local features, resulting in the low accuracy. To overcome this problem, a novel method of bundling features with multiple segmentations was proposed. In our method, images were multiple segmented, and all segmentations were described by SIFT features fallen inside the area to generate bundling features. The bundling features were matched based on the VW vocabulary. An improved similarity metric was presented to measure the similarity between matched bundling features, and the degree of similarity was infused into the vector space model of VW method, to calculate and sort the image similarity. In the process of SIFT features extraction, the number of SIFT features extracted from an image is usually large and cannot be precisely controlled, which results in poor system performance for low and unstable efficiency. A SIFT pruning algorithm is proposed to address the above issues. The algorithm measured discriminative power of keypoints by weighted combining of their contrast and ratio of principal curvature, and extracted the proper number of most significant keypoints within a given threshold by a two-phase filter process in the steps of keypoint localization and orientation assignment.Experiments show that the proposed method can exploit the space relationships among local features, and improve the retrieval accuracy greatly with no significant reduction in the efficiency. And with the SIFT pruning algorithm, the system efficiency and stability are further improved.