The Realization Of High-precision Image Classifier Based On Feature Subspace

The image semantics retrieval and the classification were recent years's research hot spots, was an important component of multimedia information retrieval, and received more and more widespread attention. Because the image semantics has the complexity, in the extraction, the expression and the application is quite difficult, therefore, the image semantics's retrieval and classifies become challenging extremely research subject. There are two big difficult problems in image retrieval based on the text method early time: First, to carry on the artificial labelling project to the image is vast; Second, artificial labelling has the subjectivity and the inaccuracy. Therefore has developed in the 1990s based on the content image retrieval technology, through the automatic extraction image vision feature, and carries on the similar match, obtains the retrieval result. This method abandoned has needed to carry on artificial labelling based on the text retrieval to each image the arduous work, has avoided the artificial labelling subjectivity. This technology obtained the rapid development, becomes the image retrieval domain gradually the mainstream technology, and has developed the massive retrieval system.But, because the image low lever feature and between humanity's understanding has the very big difference, the image contains the semantic content is unable with the image low lever feature to come the accurate indication, namely, in the image first floor vision feature and the image contain between the semantics has"the semantic gap". In view of this question, proposed the image semantics retrieval and the classified technology. The image semantics retrieval and the classified technology unify the semantic information and the low lever feature, carries on the retrieval and the classification to the image. this technology's key point lie in how to withdraw, to express, as well as use semantic information. This paper has carried on the discussion to the image semantics retrieval and classified topic's certain questions, including the image semantic extraction, the semantics indicated that as well as image semantics retrieval and classified technology.The first chapter elaborated this paper background knowledge, including the semantics image retrieval and the image classification's related content, simultaneously in detail showed the semantics image retrieval overall frame and this paper in the semantics image retrieval position. The second chapter explained the sorting algorithm essence and the type, explained the machine learning developing process and the commonly used algorithm. The third chapter concerns this paper to realize sorting algorithm foundation - support vector machines (SVM), narrated the support vector machines theory origin, the core thought and the commonly used nuclear function. The fourth chapter proposed this paper image sorting algorithm, namely based on the feature subspace's image classifier, unified linear nuclear function SVM and the AdaBoost taxonomic approach, obtains one kind of comprehensive classifier, the AdaBoost-SVM two-class classfication mean accuracies which the experiment obtained achieves 94%, higher than other taxonomic approach far. Afterward proposed the feature subspace method, fuses well the different feature and realizes the image classification. The fifth chapter the classifier which proposed to this paper has carried on the accuracy test, altogether has performed two experiments, respectively be image classifier multi-classification accuracy test and feature subspace classifier accuracy test. The experiment counts obtains the classifier precision is 84% finally, is higher than at least 10% compared to the sole classifier method, compared to ordinary and so on weight feature space classifier classification precisions is higher than 5%.The main work completed by this paper as follows:(1)This work has realized two kind of classifier SVM and AdaBoost unifies, takes Adabooat SVM the weak classifier, the weak classifier has chosen linear kernel SVM, this kind of SVM operating speed is quick, the parameter is simple, easy to operate, more importantly when with the AdaBoost union, does not need the weak classifier itself to get very high precision, the iterative process is in itself increases the classifier precision the method. In the experiment we discovered SVM classifier when processing classification question the effect basic compares the doctrine of the mean, but the erroneous fluctuation is small, the classified precision maintains stabler, this is the SVM feature. But AdaBoost in processing simple big sample multi-tag sort question time, is merely has the high precision in the training process, but once has met the strange sample, in the experiment manifests specifically when the test sample, the AdaBoost classifier will appear expected, but the study phenomenon really, this kind of fatal defect has caused in the processing image classification time it will be unable to apply directly, because practical application Chinese Library likely retrieving itself impossible all retrieval image for known, meets when the unknown sample will examine the classified duty, the AdaBoost classifier appears too many problems.(2) In AdaBoost carries on iterative to the SVM in the process, it is used the parameter value which changes unceasingly according to the circulation.because AdaBoost classifier in classification time requests the diffience between two neighboring weak classifiers bigger than certain value , if the neighboring weak classifier difference cannot meet the requirements, then the weak classifier will think that the classifier already did not have the precision promotion possibility to jump out the circulation. Therefore, we design classifier's time, has not requested the internal weak classifier is the linear kernel SVM classifier fixes its internal parameter. But linear kernel SVM itself has a C parameter, may the very convenient adjustment each time iterate the C parameter achieves increases each time the weak classifier different goal to realize the iterative process optimization. We establish a scope, makes each time in the iterative process this C parameter stochastically to change.(3) the image classification Carried on in the characteristic subspace.And it will possess the eigen vector/feature vector/proper vector each univariate to take a sub-space and to entrust with the weight, simultaneously in the iterative process, each time the iterative process carries on the coefficient weight to it the adjustment, unified the AdaBoost itself inherent sample weight and the weak classifier weight constituted the AdaBoost-SVM classifier three weighting factors, through adjusted these three coefficients to constitution the final strong classifier.In the end of this paper it talks about the place which needed to improve, and it has analyzed that it should unify the region splitting technology in the image semantics classification and the retrieval. Simultaneously this article has carried on the forecast to the next image classification's development, proposed that may unify the machine learning with the multi-demonstrations forms the synthesis study method. Also proposed that the feature extraction process's importance, may the feature fitting technology which proposed newly apply the classification learning process.