Research On CBIR Test Platform And Related Techniques

With the rapid development of network technique and multimedia technique, there is much more information on the network and multimedia information has been used much more widely. Image information is the largest and most important of multimedia information. Image retrieval is the key and hot research of current multi-media retrieval and image technique. CBIR (Content-Based Image Retrieval) retrieve using the feature of image which is automatic extracted by machine can largely make up the insurmountable flaws and shortcomings of traditional TBIR (Text-Based Image Retrieval).From the rise of CBIR to now, CBIR algorithm researchers have proposed many different algorithms. However, these algorithms use different validate methods and evaluate criteria, therefore their results cannot be compared objectively between each other. How to validate and evaluate them using a unified and objective way is the key. Build an algorithm test platform which use unified and objective validate method and evaluate criteria not only can improve the development of CBIR algorithms but also can promote the establishment of reliable CBIR commercial applications. This paper design and implement an efficient, easy to use, scalable, web-based CBIR algorithm test platform which can be used for algorithm test and evaluate. This test platform uses modular and hierarchical architecture design methods. It has a simple and efficient interactive interface. Its modules are powerful and be decomposed clearly. It provides both subjective and objective evaluations of search results.Index structure is one of most important factors of which affecting the performance of CBIR system. The image feature which automatic extracted has characteristics of large storage space required, high dimension and so on. Many index structures have a pitfall of encounter"Curse of Dimensionality". In this way, index technique has been the key problem which impacts the progress of CBIR. Base on in-depth analyzing the classical tree index structure and VA-File index structure, this paper proposed a two-level vector Approximated, aggregative and sectional index structure. This index structure is based on VA-File index structure. Experiments show that this index structure not only conserves the good properties of VA-File, but also can improve the retrieval performance better and simplify operations. In this way, it has some value and significance.Single type of feature can hardly describe the abundant image information completely. Any kind of image features using in image retrieval has its advantages and disadvantage. Different features can be combined with others in order to improve the retrieval effectiveness. Based on in-depth analyzing of different levels of multi-feature fusion, this paper proposed a new method which is dynamic, easy to extend and operate simply. This method is based on MAMD (Multiple Attribute Decision Making). By using this method to retrieval, it not only can improve the effectiveness of CBIR but also provides the algorithm researcher a way of real-time fusion and observe the effectiveness of fusion retrieval When it running in Kapoix.