Study Of Image Active Classification Method Based On Resampling Thought

Image classification is an important technology in the fields of computer vision andimage processing. With the rapid development of information technology and the explosiveincrease of image data, how to select small amount of high-quality image data rapidly andefficiently in the process of learning classifier has become a research focus of the imageclassification methods.In this paper, we investigated the technology which selected training samples used fortrain classifier in the image classification, and proposed a number of relevant algorithmsand technical schemes. The main studies are as follows:(1) The sampling strategies of image classification based on the active learningalgorithm were discussed. The uncertainty sampling strategy which is widely used in theimage active classification was studied to provide a theoretical basis for the image activeclassification method based on resampling thought.(2) Both the current and expected values of the samples were considered and aMaximum Classification Optimization (MCO) algorithm was proposed to select thesamples with high value and optimize the classifier performance, according to the problemthat only current value of unlabeled samples is used in the traditional image classification.(3) Samples are selected directly from the unlabeled sample set and thus the influenceof noise samples is usually ignored in the traditional image classification methods atpresent. Accordingly, a construction method of the undirected graph model based on thelocal density information was proposed. The noise image data was processed by using theundirected graph model to reduce the influence of noise image data on the property ofsampling strategy.(4) In view of the problem that the representativeness of samples selected by thetraditional sampling strategies is poor, a new method for selecting the representativesamples based on the undirected graph model was proposed. Representative sample withhigh information was selected and labeled by combining the MCO method with the samplerepresentativeness calculated by the undirected graph model to update the classifier.In addition, the above proposed method was demonstrated by the experiments, whichindicated that the proposed scheme is feasible and effective.