Research And Implementation Of Image Retrieval Based On Visual Features

With the development of multimedia technology and Internet network, image data expand sharply. It is already a very urgent problem that how to make full use of existing data and keep them away from sinking in the information swamp. The information needed should be found accurately and quickly. The development of image retrieval can be divided into two generations: text-based image retrieval and content-based image retrieval. Content-based image retrieval has become a current hotspot because it measures the visual similarity and can avoid some shortage of text-based image retrieval, such as overloaded work of comments, too richness and subjective of image information, and so on. The image features in broad sense include:①text features, such as file names, notes, etc.;②visual features, such as color, texture, shape, and so on;③semantic features, such as objective description of image content, subjective perception of human observation and so on. Text-based image retrieval requires annotations of text features. Content-based image retrieval automatically extracts visual features or semantic features. Most of the current ways to determine the similarity between images employ objective visual features, such as color, texture, shape, and so on.This dissertation mainly focuses on research of features extraction which is a key technology in content-based image retrieval. A variety of typical retrieval algorithms have been analyzed and compared. In addition, I have developed an image retrieval prototype system.The main contribution of my work includes the following four aspects:1. This paper makes some research on various color retrieval algorithms and the structure of a color description. A new algorithm of scalable spatial dominant color has been designed on the base of the above-mentioned work. The main thinking is as follows:(1) The algorithm can reduce feature data and act in accord with human visual characteristics. The experiment has achieved relatively good results;(2) The amount of feature data and similarity calculation can be adjusted according to applications and demands.(3) The algorithm also reckons spatial information, amending the deficiency of traditional histogram. 2. There are some theoretical and experimental data to prove that the first-order moment of histogram do not have the ability to distinguish between images. The similarity function depresses the distinguish ability because it calculates with the first-order moment, the second-order moment and the third-order moment;3. Some typical algorithms based on other visual features have been researched and analyzed, such as texture, shape and so on;4. A content-based image retrieval prototype system has been designed and developed to provide the basis for related research and the expansibility for follow-up work with VC 2003 and SQL Server 2000.In this paper, I have conducted a preliminary research on content-based image retrieval, evaluated a variety of classic retrieval algorithms, put forward my own views, proposed a new algorithm of scalable spatial dominant color, and developed an experimental prototype system for the research work.