Semi-supervised Based Relevance Feedback And Its Application In Image Retrieval

The purpose of the image retrievement system establishment is the accurate retrieval results be presented effectively and fast, the reasons which affect the accuracy of the retrieve result are as follows, the machine vision algorithms cannot bridge the gap between low-level visionary information and high-level semantic information, and hence it cause the low accuracy of the result. To solve the problem, this paper did research on semi-supervised learning in the field of machine learning and relevance feedback technique in the field of image retrievement. Using the data resources fully to enforce the algorithm model's learning ability to the data distribution and retrieve intention, so that it can combine vision information with retrieve intention, and lift up the retrieval accuracy. The content of this paper describes as follows:First, this paper describes the image retrieval researching background, and overviews semi-supervised learning methods and relevance feedback, analyses the problems in the retrieval, and makes a description of the semi-supervised algorithms. It points out the situations that retrieval system be faced with huge data.Then, this paper proposes a semi-supervised method based on kernel-ELM in chapter 3. This method makes expansion on the traditional co-training method, as framework of tri-training, the kernel-ELM brought in this framework, makes it as the elementary classifiers, using data fully enhances the sensitivity of classifiers by combining the neural learning method and the standard optimization. The experiments has tested on the standard public dataset on semi-supervised algorithms. The experimental result shows this method has strong classification ability with high robustness and generalization ability.Afterwards, this paper proposes a method of relevance feedback based on kernelELM semi-supervised algorithm in chapter 4. After combining the semi-supervised method, and make it fit into the relevance feedback. This paper performances the experiment on public dataset. The method can performance with high accuracy and low time-consuming because the characteristics brought by the semi-supervised algorithm and the timing advantages owning by the kernel-ELM. Furthermore, it reaches the purpose of relevance feedback.Last, this paper aims at the actual bottleneck of the retrieval facing a bunch of high-dimension features, and optimizes the time-consuming problem of traditional retrievement system. Based on the foundation of fusion-featured “Bag Of Feature” retrieval model, the Inverted-file-index technique has been led into the description of image content. It optimizes the Curse of Dimensionality during the procedure of similarity measurement.