Research On Content-based Image Retrieval For Techniques And Application

With the development of multimedia technology, there comes out a mass of multimedia information. How to retrieval the information becomes a research focus. The key technologies of content-based image retrieval (CBIR) system contain a lot of aspects. The most important point is how to represent multimedia content accurately and completely. The establishment of MPEG-7 Standard gives a direction of the standardization of multimedia content representation.First, the paper discusses the most popular CBIR technologies mainly based on low-level descriptor such as color, texture and shape. Then we develop a plan to evaluate the typical technologies of CBIR. We analyze the typical technologies throughout the plan laterally and vertically; and point out their performance and sphere of application.Second, we introduce the MPEG-7 standard, discuss the content and application sphere of the MPEG-7 standard, and analyze the practicality of assembling MPEG-7 standard into our CBIR system.And then, based on the evaluation result of low-level descriptor retrieval system and the complexity of implement of assembling MPEG-7 standard into the CBIR system, we use the Scalable Color descriptor Color Layout descriptor and Edge Histogram descriptor to develop the CBIR system. We improve the system by combining the three descriptors. The focal point of our work is how to represent the image content: dividing the entire feature of image content into three-tier structure. Our work also focus on the representation and extraction of the low-level color and texture descriptors, the similarity measurement of single feature, and the normalization of similarity measurement of combined features. Our work also includes the overall design of the system framework and the design of high-level semantic retrieval module.Finally, the paper discusses the implement of the CBIR system, analyzes the runtime performance and reliability of the CBIR system. We contrast the implement of using single descriptor and the implement of using all descriptors. It turns out that the performance of integration of three descriptors is better than the performance of using single descriptor.