Research On Some Similarity Metric Algorithms In Image Retrival

Gaussian Mixtures(GMs) method is a popular approach to represent the content of an image. It carries more imformation than histogram method. There are lots of similarity metric based on GMs. Tow similarity metrices are descussed in this dissertation, the first one is optimization method, typically Earth Mover's Distance (EMD); the second one is probabilistic approach, typically asymptotic likelihood approximation (ALA) distance. This dissertation also research on the methods that fit GMs model. The main works of this dissertation are summarized as follows.Firstly, EMD conducts a similarity metric measurement by minimizing costs, which performs better retrieval result with relatively low computational complexity. The accurate measurement benefits from the properly formulated probabilistic model of an image, while the models based on histogram and vector quantization (VQ) cannot efficiently describe the image information. In this paper, we use Gaussian Mixture EM (GMEM) algorithm to describe an image and form a better probabilistic model for EMD algorithm, which results in the GMEM + EMD algorithm. Experiments show that the EMD measurements based on GMEM clustering can improve the retrieval performance effectively. We also contribute a new feature representation method, which increases the weights of image textures, reduces the number of the subblocks of an image, and thereby speeds up the computation.Secondly, Similarity metric is a key step of content based image retrieval (CBIR) and it also one of the difficulties. Probabilistic approaches are a promising solution to the image retrieval problem, and ALA algorithm is one of the excellent measurements of probability similarity function, when the mixture component of the query spread around the experiment results show that the ALA algorithm is the best. Because of it's easy for this algorithm gets bad performance; we use multi-layer approach to reduce the influence of this situation, and call it Improved ALA (IALA) algorithm.Thirdly, if the mixture component of the query group together, the ALA algorithm performs badly, this is the time we use EMD as similarity Metric instead, and we call it measurement selecting (MS) algorithm. It uses the probabilty feature of query's GMs to select IALA or EMD as similarity metric of CBIR system.Finally, CBIR is mainly contained of tow phases: first, to represent an image; second, to measure the dissimilarity between images. In this paper, we introduce an improved expectation-maximization (EM) algorithm for clustering GMs to represent an image, instead of regular EM algorithm.