Entropy Based Image Retrieval

With a good deal using of image, it is more important to image information resource management and retrieving. The principal research of content based image retrieve (CBIR) includes two aspects: visual feature representation and similarity measurement. In general, visual feature representation has three levels: the image level, the vector level and the number level. The majority representations of visual feature are the vector level. The space of similarity measurement is N dimensions and the calculation is very large.For aggregating the level of visual object to the number level, we propose one method : entropy as a visual feature of images. Thus, the space of similarity measurement may be decreased to one dimension and the speed of retrieving can be improved. With the color histograms as the probability density function of images, we can define the entropy of images.With the concept of entropy as the visual feature representation , we propose two methods of similarity measurement: entropy difference distance measure(ED) and maximum relative entropy distance measure(MRE) . ED is the simple subtraction of image entropy, so the speed is fast; the relative entropy does not satisfy the symmetry axiom , and so we propose MRE. Since MRE is not a metric in the strict sense of the word, we use ED as the similarity measurement of entropy.For two very dissimilar images , their entropy may be equal. So only using ED to retrieve images is not proper and exact. For undertaking the fast speed and exactness of the retrieving algorithms , we propose a new retrieving algorithms : EDLN . The entropy difference formula can be applied to the database to return an initial set of retrieved images. This initial set, which is smaller in size than the entire database , is then searched using the Lrnorm to retrieve a final set of images similar to the initial query image. Based on EDLN, we develop a set of system of retrieving flags with VC++ 6. 0.In a word, this thesis principally includes:1. Entropy as a visual feature of images. The level of visualobjects of images is aggregated to the number level from the vector level. The space of similarity measurement may be decreased to one dimension.2. We propose two methods of similarity measurement: entropy difference distance measure(ED) and maximum relative entropy distance measure(MRE) .3. we propose a new retrieving algorithms : EDLN . EDLN undertakes the fast speed and exactness of retrieving . Based on EDLN, we develop a set of system of retrieving flags, and EDLN can be applied to retrieve images with color feature being obvious ,such as trademark , logo and so on.