Research Of Image Representation Methods For Image Retrieval

In recent years, more and more data aregenerated every day. How to effectively retrieval the required information from the stored data becomes the key issue of information management system. As images occupy a high percentage of the avaiable data, image retrieval is a vital part of information retrieval. Through image retrieval, we want to find the images which are similar to a query image.Image retrieval involves two important tasks, including image representation and data retrieval. The results returned by an image retrieval system strongly depend on image representation methods. Image representation methods have significant effects on the efficiency of image retrieval system. In this thesis, we focus on the image representation methods.Image representation methods can be divided into two classes, including local feature based methods and global feature based methods. This thesis analyzes both local feature based methods and global feature based methods, and proposes two novel local feature based image representation algorithms. The main work and contributions of this thesis can be summarized as follows.1) It proposes a novel feature weighting algorithm based on the dissimilarity of original local descriptors. In image retrieval, repeated visual elements usually appear in an image, which will significantly affect the system performance. This phenomenon is called burstiness. In this thesis, we exploit the dissimilarities between different SIFT descriptors to compute weights of the descriptors of a query image. These weights are used to reduce the influence of intra burstiness. To reduce the compuational complexity, global saliency is adopted and intergrated with our feature weighting algorithm. The feature weighting algorithm can improve the precision of image retrieval with a small increase of time cost.2) It proposes a novel query pruning algorithm. In image retrieval, features are extracted from the query image. Some features are informative while others are uninformative. The uninformative features will reduce the precision of image retrieval. We propose a method to compute a numerical value that evaluates the importance of a feature. The method takes the retrieval results of a natural image database as a reference signal, and the retrieval results of the main database as another reference signal. The features are sorted according to their numerical values. Some features in the query image will be pruned, which can improve and the precision of image retrieval.3) Expriments of the Convolutional Neural Networks features are conducted. Not only the performance of several network architectures, but also functions of different layers of Convolutional Neural Networks are investigated. The functionality of the fully connected layers is analyzed by considering a multi-scale and multi-location method. On the end, the data retrieval algorithm is taken into the system to yield desired performance.