Vocabulary Tree-based Image Retrieval Technology Research

The development of information technology gives an impetus to the development of multimedia technology, along with the development of information technology and the multimedia technology, image information is also growing rapidly. People began to care about how to management these images scientifically, legitimately and effectively, how to identify needed images quickly and accurately from the huge image database. Therefore, Image retrieval technology becomes a hot research direction.The traditional content-based image retrieval technology expresses the content features of each image by vectors, stores the large numbers of feature information in the database. via comparing the similarity of eigenvectors of images to search and images in the database, and through calculation, output several images according to the descending order of similarity, then the image retrieval process is completed. This method shows good accuracy and high efficiency when the image database is not very big. However, when more images, this method will not satisfy the requirements of users for retrieval time and accuracy. This paper based on the Vocabulary tree image retrieval method, it is a kind of new content-based image retrieval method. This method can effectively deal with the challenges which the traditional content-based image retrieval faced when the image databases is bigger.Firstly, the key technologies of content-based image retrieval were studied, the extraction methods of image characteristics were analyzed. The SIFT eigenvector which is invariant to the color, rotation, and translation,etc was chosen as the content feature to depict image.Secondly, adopted the method of hierarchical k-means clustering, the extracted image features were generated to visual vocabulary. According to the train of thought of key term TF-IDF in text retrieval, the generated image visual vocabulary was related by TF-IDF, then, the vocabulary tree of eigenvectors was constructed, the generated process of vocabulary tree was simulated by software. Owing to that the feature vectors in the feature library were constructed to vocabulary tree, and the prior node and the following node of tree structure have the relation of inheritance, the image retrieval based on vocabulary tree saves time for image matching, especially for a larger image library, this kind of advantage is more prominent.Finally, the image retrieval method based on vocabulary tree, in which the image's content features were extracted by 36 dimension PCA-SIFT features instead of 128 dimension SIFT features, was put forward to reduce the computational cost of computer. The image retrieval system based on vocabulary was realized by software, visual man-machine interactive platform was designed. Using the images in Corel image collections for retrieval experiment. The experiment results show that using PCA-SIFT algorithm which has reasonable dimension to extract the content features of images for image retrieval based on vocabulary tree, compared with the SIFT algorithm, reduces the retrieval time obviously on the premise of ensuring the retrieval accuracy as far as possible.