Image Classification With Shape Information

In the face of massive image data resources on the internetwork or websites, how to efficiently classify big image data resources and manage the mass image data appropriately. In recent years, it has become the urgent problem needs to solved in the field of computer vision and pattern recognition. Image classification technology based on semantic is able to perform classifying image data resources automatically by using the characteristics of the image itself, is the best way to solve the problem of mass image classification. This paper aims to study the automatic image classification, and it has important research significance and application value. The main work of this paper includes:The first work is the image classification based on regions non-uniform spatial sampling. Firstly, the over-segmentation technology is performed on the original image, then, it could obtain the over-segmentation regions of the image, and utilizing saliency detection method to estimate the importance of each pixel in the over-segmented regions. Second, without increasing the sampling number of local features, dense uniform sampling is applied to the important boundary of salient region, random sampling inside the salient region according to the area size and importance of regions. Finally, the bag of words representation model is used to accomplish big image data resources classification.The other work is the image classification based on image cluster multiple kernel learning model, which introducing the intermediate representation layer between the object class and individual image--" image cluster ", to catch the characteristics of sample image as well as to maintain the stability for each category in training classifiers simultaneously. Given a target object class, the image samples which have similar characteristic of the visual appearance will be clustering into the same image cluster, therefore, it could control intra class differences through a series of image clusters. On the other hand, inter class correlation be able to control through the relationship between images of image cluster belong to one category between images of image clusters belong to other categories. So, seek reasonably coordination between inter class correlation and intra class difference, can effectively handle image data resources classification problems with similar scenes.Experimental verification are all performed on the widely application datasets (UIUC Sports Scenel5, Caltech101/256, Pascal VOC2007). The experimental results show that proposed local feature extraction method which based on local regional non uniform spatial sampling strategy is able to further improve the image classification accuracy of dense sampling local feature extraction method. In addition, the experimental results of based on image cluster optimization multiple kernel learning model prove that, image cluster optimization multiple kernel learning model can handle the inter class correlation and intra class difference of train samples, and effectively improves the image classification accuracy of the multiple kernel learning model.