Scene Images Categorization Based On Codebook Model

Natural scenes classification is an important challenging task in computer vision.For classifying scene images automatically using computer, the most critical issue is toextract distinctive features from images and form distinctive image representation. In thispaper, we focus on the methods of features extraction and image representation.First of all, local features are commonly used in scene classification for featuresextraction. However, the different features have contribution on different degree. In orderto highlight the significance of salient features, this paper classifies the local featuresbefore codebook learning, and improve the discriminate of image representation viacontrolling the codebook size.Secondly, the structural information is the important information of images whichplay an important role in natural scene classification task. This paper adds structuralinformation for images applying the technology of Superpixel Lattices segmentation. Thetechnology improves the shortcoming involved in the regular rectangular segmentation.Each sub-region obtained using Superpixel Lattices segmentation technology hasintegrality and the features in the same sub-region have consistency. In addition, the imagelocal features are re-constructed contextual features to add local structural information forimages representation.Finally, global features are sufficient for separating scenes with significantdifferences in the global properties, while the local features are suited for the images withsimilar global properties but local differences. Furthermore, the extended contextualfeatures provide more comprehensive contextual information. Also, this paper improvesthe current features mapping method LLC(Local-constrained Linear Coding) for theregion representation. Furthermore, global and local features are weighted combined topromote them can have complementary advantages. The image representation based onthis technology will be more comprehensive and more distinctive.This paper mainly do contribution on three aspects: features extraction, features mapping and images representation. We set15-category scenes as our experimentaldatasets and apply libSVM with HIK(histogram intersection kernel) to complete scenecategorization. The ideal experimental results illustrate the fine effectiveness of ourapproach.