Research On Compact Feature Based Clustering For Large Scale Image Retrieval

With the popularization and development of the Internet and digital photography equipment, the number of images become large and the content is more and more rich.The image serves as a content-rich intuitive digital data form, and its powerful advantage of information make it rapidly permeate the society from all walks of life, with increasing applications. For the image data of explosive growth, it becomes particularly important that how to retrieve the interested images.In recent years, the researchers focus on the content-based image retrieval technology. They take the theories of computer vision, information retrieval, machine learning as the foundation, and develop the technologies of image visual content extraction and retrieval, and have made certain progress. When facing the growing massive image,real-time demand for image retrieval becomes increasingly important. However, the traditional content-based image retrieval can hardly meet the requirement. To address the shortcomings of the existing methods, we tudy the start-of-the-art achievements in computer vision, information retrieval and machine learning, and do in-depth researches on the visual feature extraction, image representation, and index.For the dimension disaster caused by the large number of high-dimensional local features in images, a simple and effective framework of image local feature aggregation is proposed in this thesis. The local features in an image are aggregated into a image representation, thus it avoids the computation overhead resulting from the quantization and retrieval of a large amount of high-dimensional features, reduces the hardware requirements, and improves the efficiency of feature description and retrieval. According to the clustering objects, the image representation is divided into two types, compact feature based clustering for image and compact feature based clustering for image database.The first clustering feature clusters the local features of an image, extracts the main descriptive information of each cluster, and computes the distribution of local features and spatial statistic information as supplementary. Meanwhile, on the basis of building index using single visual information, we propose a retrieval method based on vocabulary tree, which utilizes multifarious information. The combination of the proposed retrieval method and the compact feature based clustering for image can significantly improve the efficiency of large scale image retrieval, while sacrificing certain retrieval accuracy.The second clustering feature first clusters the local features extracted from image database, splits the whole local feature space, and then assigns the local feature into each partition space to form the compact histogram, according to the correlation of distance and spatial information between the local features and the clustering centers. The compact feature based clustering for image database takes advantages of the global characteristics and statistic information of the local descriptors in the image dataset, and records the distribution of the local descriptors of each image relative to the entire local descriptor set, thus the descriptive ability and discriminative power of each image descriptor is improved.To sum up, a large scale image retrieval framework using compact feature based clustering is proposed. The thesis designs the retrieval schemes for the two compact features, and verifies the effectiveness of the proposed features by conducting extensive experiments. The experimental results demonstrate that the compact features outperform in terms of scalability, and can apply to large scale image retrieval.