Classifying Images Automatically Using Hypersurface

In the society, image files have been increasing at unbelievably rapid speed, how to classify and manage these files is on people's focus more and more. The application background is also very broad, prohibition of unhealthy images, automatic management of massive images, and so on.Aiming at retrieving and automatically classifying of a large quantity of images, the influence of image size on the efficiency of image processing is of essential importance. For promoting efficiency, the concept of interest block is proposed in pretreatment of image segmentation, and that is the key to get a miniature of the original image. And then we segment the image into several continuous blocks, we suppose that the probability of these blocks being single objects is high, so firstly we extract the feature of blocks, then project them into the feature space, and the key to identify singular objects is training and classifying these singular objects using improved hypersurface method. At the beginning, the number of features to be extracted is big, and that means the dimension of feature vector is large. For reducing the dimensions of feature vector, we eliminate some features which contribute little for correctly classifying one object from another or have the same effect as other features using data mining. To bridge the gap between the low-level image features like color, texture which are used for traditional image classification and the high-level concepts, hierarchical image classification framework is betaken. On the base of identifying the single objects of an image, the whole image is then classified.Our experiments on Corel image library and LabelMe image library which is popular in research obtained very positive results not only in high correct rate, but also in high efficiency processing massive images.