Research On Image Classification Algorithms Based On The Transductive Multi-instance Learnig

With the rapid development of the multimedia and Internet technologies, as well as the popularization of the digital products, the number of all kinds of digital image increases explosively. Hence, how to manage and apply these digital images to every field effectively has become a new research hotspot, where classifying is one of the urgent issues. The traditional method of image classification is generally based on the images that have been labeled manually. However, there are two intractable problems. Firstly, the effectiveness of the method is restricted by the human itself. In another words, manual annotation on images is often susceptible to intensive subjectivity. Secondly, manual annotation is too time-consuming and arduous to apply to a large number of images. Research on content-based image classification starts from the90’s of the last century. CBIC classifies images by processing and learning from the extracted the low-level features. There have been great achievements in CBIC, and only the one single feature is generally used in the methods. Since there is more than one object in an image, it is not enough to use one feature to describe the image. The method of multi-instance learning (MIL) can deal with the above problem. By intensively studying MIL and support vector machine (SVM), we proposed two new MIL methods to classify images.The main contributions of this paper are as follows:1. Based on transductive support vector machine (TSVM), we provide an MIL algorithm (DD-TSVM). First, the diverse density algorithm (DD) is used to find the local optimization points in the instance space, by which the feature space is constructed. Then the bags are nonlinearly mapped into the feature space. Finally TSVM is used to train the classifier. The proposed algorithm effectively takes advantage of the unlabelled samples. The experimental results on Corel dataset show that DD-TSVM algorithm has good performance.2. Aiming at the redundant data existed in the training data; we provide a MIL algorithm combined with feature reduction (DDRS-TSVM). A rough set based on neighborhood is incorporated in DD-TSVM algorithm to manipulate the MIL training data, which eliminates the influence of redundant data on classification. The experimental results on Corel dataset demonstrate the performance of DDRS-TSVM, which outperforms DD-TSVM.