On Content Representation And Catigorization Of Scene Images

With fast development of image sensor, network and storage, number of images in offline or online database has being increasing rapidly. It is a tremendous challenge to manage effectively and retrial images rapidly. Appropriately describing and classifying scene images are two items of the key technology to solve this problem. Based on the past study of scene classification with features, methods with visual words have occurred and become popular. This thesis focuses on image features extraction, to image scene representation with visual words and image classification with fusion of multi-type image features.it is essential to explore simple and effective new texture features since there are large amount of texture contents in scene images. We propose a new texture feature based on multiresolution histogram moments. It not only describes effectively the spatial variation of texture which resulting in decrease of feature dimension, but also is more robust to noise. This new feature is powful support for the following scene classification study.Based on the visual word representation of images, two aspects of studies are proceeded: the multi-level topic model of image content and the model based on context information of visual words. The former extends classic LDA and pLSA model and utilizes multi-scale and multi-type feature of images, and two multi-level topic models of scene images are established, respectively, namely MT-pLSA(Multiple-level Topic pLSA)and MT-LDA(Multiple-level Topic LDA). These models improve classification performance.The later uses Spatial Pyramid matching to model the position information of visual word pairs and visual word groups, and gets the context pyramid features. With these features and SVM classifier, better scene classification performance is obtained.To bring to play the potential of scene content representation based on image features, we study the method which fuses multiple-type features.Multi-type features of images include: spectrum features and spatial-spectrum features based on sub-blocks of images; global spatial features and spectrum features of the whole images. The fusion method is to form a progressive scene classification framework by using stacked-SVM.This method is simple and stable and gets the classification performance which is better than most popular methods with visual words.