Visual Vocabulary For Image Retrieval Tree Study

As the development of technologies related with multimedia, the scale of digital image has expanded rapidly, as well as the application of that has become more and more widely. Therefore, how to retrieve the needed image from a large-scale image database effectively and quickly has become a hot topic. Because of the disadvantages of the traditional image retrieval techniques which are based on text description and keyword, such as the subjectivity and the inaccurateness caused by artificial image tagging and so on, they cannot satisfy the requirements of users for inquires. The main content-based image retrieval technique is based on the bottom visual feature, the retrieval system extract the visual feather from images directly. The images from the database and the inquire images have been expressed as feature vector, and the systems calculate the similarity of feature vector to select the retrieval results. Those techniques have a good performance on image retrieval.In this thesis, given a large number of features extracted from images, we construct a vocabulary tree by using the algorithm of hierarchical k-means cluster. There are two parameters:the number of levels L specifies the height of vocabulary tree, and the branch factor B specifies the number of children each node has. Vocabulary tree is a tree structure of features organization mode, each node of the tree called a visual word. The features extracted from images are quantified as a visual word by using the vocabulary tree, thus an image can be pressed as a discrete visual word vector. In the vocabulary tree, as same as the form of text retrieval, each visual word in the tree is associated with an inverted file, then calculate the distance between the visual word vector to measure the image similarity. By introducing the concept of vocabulary forest, we can realize the precise division for feature space. Through the hierarchical quantification of image features, we construct a multi-resolution visual word histogram. The method of Pyramid Matching Kernel can realize the rapid matching of feature vectors, which is suitable for large-scale image retrieval system. In order to adapt to dynamic image retrieval system, this thesis also studied that when image database changes, the corresponding vocabulary tree adjust its structure correspondingly according to its update algorithm, to adapt to the changing of image database. In addition, through the experiment, we tested the effects of growth parameters on retrieval performance.