Image Retrieval System Based On Parametric-statistics

With the development of the information technology, it is more convenient toaccess a camera equipment and take photos, and the number of pictures on theInternet explodes,so there’s more and more demands for image retrieval.However,traditional retrieval is based on text retrieval,as the “semantic gap” between textualand visual features exists, it’s far from perfect for the traditional method. There’s agreat development for the image retrieval based on context.Bag of Visual Word model makes the image retrieval for a large scale databasepossible, however, there are still works should be done. First, the discriminativepower of visual features decreased rapidly when quantized into visual words. Second,the basic BOV model regard each descriptor as a unit which ignore the contextinformation.Our works are mainly on these two respects, one is using hamming codeand soft quantization to reduce the loss of quantization which should significantlyincrease the discriminative power of descriptors. We then introduce the distance of<mean, variance>, on the basis of which we construct the old visual codebookgenerated by K-means as a undirected graph. Then we apply the adjusted Prim’salgorithm to calculate the minimum spanning tree,the new order of the outputsequence of each visual word is the new visual codebook order.On the next stage, foreach Hessian-Affine context region,we calculate the parametric statistics such as areaof the region, density,scale,angle and variance of visual words’ IDs.When matching,we take the parametric statistics into account, re-evaluate the weight according to theconsistency between query region and the region from the database image. By doingthis, we expect to eliminate the mis-matches. In our work, we realize a image retrievalsystem for experiments, and test and verify our method on Holidays and UKBench.which proves our method significantly improved the retrieval performance of BOVmodel within real-time response.