Study On CBIR Algorithm Based On Color Cluster And Spatial Relationship Of The Object Regions

Along with the development of the multi-media technology, image information emerges in abundance. How to effectively manage and use these resources has always been the focus of the scientific studies home and abroad. Traditional text-based image retrieval (TBIR) can no more satisfy the current needs. Effective obtain and use of information can only realized by resorting to content-based image retrieval (CBIR). Therefore, various CBIR algorithms have been put forward. Based on the former studies, this thesis puts forward a new CBIR algorithm based on the color cluster of the object regions and the spatial relationship, which is proved effective in the tests.The thesis covers the following aspects:1. Introduce the historical background of CBIR and its conceptual model, especially the key technology related to this thesis.2. Extract the object regions from the image by using the improved algorithm. In this thesis we extract the edge information from the luminance image by using Canny algorithm, and connect these edges. We select a maximal adjacent connection aggregate from the binary edge image that has minimum external rectangle area which corresponds with the minimum external rectangle as the object regions of the images.3. The initial centers of the cluster are another emphasis in this thesis. This thesis put forward the DBEC algorithm that can ascertain the initial clustering centers which used the data structure of single list and string. The result proves that the initial clustering centers are limited to a small area which can approach to the best initial clustering centers more closely.4. In this thesis we select three cluster result centers to make up the triangle which expresses the spatial relationship. This thesis puts forward the CTIA algorithm which through comparing the difference of corresponding inner angle of the triangle to be the similarity of these images.Results show that the algorithm is effective in both recall and precision. However it is insensitive to the rotation of the image and the change in size which has superior robustness. In the ending part of the thesis, a summary is made and the areas for further study and improvement concerning this algorithm are pointed out.