Natural Scene Classification Based On Visual Features And Theme Models

Image classification and management is a difficult and challenging task with the development ofmultimedia technology. In recent years, natural scene classification as an effective way for imageunderstanding becomes a research hotspot. The task of scene classification is the automatic annotation andclassification of scene images based on semantic category to provide guidance for target recognition.The work of this thesis is focused on three aspects including feature extraction, feature selection andtheme model selection, with the analysis of three layer structure model of natural scene classification. Infeature extraction, in order to get rid of the disadvantages of the traditional BoW model which lacks therepresentation of semantic space, the spatial information is generated by using the combination of globalfeatures and local features with a space pyramid method. In feature selection, a new feature selectionmethod using the Bayesian harmony is proposed. In theme model selection, SLDA model is adopted inorder to get the latent topic and classification.The main results obtained in this thesis are as follows:Scene classification method based on the specific category selection is proposed. It uses spatialpyramid model to generate low-level features, and then integrates into the BoW model. This methodacquires the visual words closest to scene category by using Bayesian harmony feature selection algorithm,further to generate the category histogram. The method retains the full spatial information, and makes astrong category presentation skills in the BoW model.The proposed scene image classification method is finally applied to the classification of water qualityimages. Water quality image classification method based on multi-feature combination is proposed.Experimental results show the proposed scene classification method for the water quality images analysis iseffective.