Research On Images Distance Metric Learning Technology

With the development of internet and computer technology in recent years, computer visiontechniques, such as image retrieval, image classification, image match and image annotation,have become research hotspot, because of demands on massive multimedia data processing. Atpresent, a common problem has to be dealt with in all these researches, which is to get thesemantic similarity through the distance metric of bottom features of image, and achieve theimage understanding. However, due to the semantic gap between low features and highsemantics, different image features demand different distance metrics, and similarity may bedifferent according to different references. Traditional distance metric approaches ignore thebackground environment and semantic restrained information, thus the result is oftenunsatisfactory. Distance metric learning brings the structure information of features space andsemantic restrained information into distance metric learned from the training datasets, and getsa distance metric function reflecting the characteristic of feature space and the relation ofsemantics.This thesis mainly researches on three problems existing in distance metric learning, that isthe problem of algorithm complexity and inconsistency, as well as the problem of its applicationsin BoVW (Bag of Visual Words)ˋthe concrete contributions are listed as follows:(1) The present distance metric learning approaches need to solve positive semi-definiteprogramming problems, and proceed with matrix derivation and matrix inversion in eachiteration, which cause the algorithm complexity of distance metric learning proportional tosquare or cube of feature’s dimension. Therefore, distance metric learning approach based oneigenvalue optimization and feature grouping is presented. The high-dimensional features issegmented into several sets of low-dimensional features according to their dependence in thenovel approach; then distance metric matrix is learned for each set of low-dimensional featuresby eigenvalue optimization distance metric learning; finally, the metric result of each set offeatures is fused together, thus, the algorithm complexity is reduced. Experiments result of imageclassification on Corel datasets show the competitive performance of the new approach.(2) The training datasets is often insufficient in distance metric learning problems, whichcauses the problem that the training datasets distribution can’t reflect the real world datasetsdistribution, and the distance metric learned from the training datasets may be not fit for realworld datasets. In order to resolve this problem, a distance metric learning approach withprobability density ratio estimation is proposed. In the novel approach, the probability densityratio of training datasets distribution to real world datasets distribution is estimated by the directprobability density ratio estimation; then the ratio is brought into Neighborhood Components Analysis, and samples are weighted with the probability density ratio, thus, the inconsistency ofdistribution between training datasets and real world datasets which existed in traditionaldistance metric learning is resolved. The classification experiments on Corel datasets and UCIdatasets show the competitive performance of the new approach.(3) Due to the weakness of feature detection, drawbacks of clustering algorithms, and thequantization error existing in BoVW methods, the visual vocabulary achieved by BoVW facesthe synonymy and polysemy of visual words, which makes it unsatisfactory to measure thedistance between images using BoVW. In addition, the complexity of learning a full matrix isunacceptable on account of huge dictionary. Therefore, a SVM based BoVW distance metriclearning approach is put forwards. A hyperplane is trained by SVM (Support Vector Machine) inthe novel approach which can extremely separates similar image pairs and dissimilar image pairs,and each dimension’s weight is got for point product of the visual words histograms. The desiredmetric matrix is a diagonal matrix composed by each weight, which measures similarity ofimages better. Experimental results of image retrieval on Oxford datasets demonstrate theeffectiveness of this approach.