| In this information explosion society, Internet has greatly facilitated people to search for information, at the same time there are also some problems. How to retrieve from a plenty of information to be useful on people’s own part involves retrieval and classification. All along image classification and labeling are the main focus areas of computer vision and machine learning, as an important means to obtain information and image semantics they has been widely used.With the advancement of technology and the popularity of cameras, more and more people uploaded pictures which they like to the network and shared them to other people. This also lead classification of natural images becoming a hot issue in recent years. Traditional image classification is usually based on artificial labeling, but there are two difficult problems:one is that labeling image manually often has highly subjectivity, different people has different label ways, it can lead to different classification results; the other problem is that many labeling projects are very huge, labeling them manually is not only time-consuming and it is very laborious, this makes it difficult to carry out in large quantities.Currently research on image classification is most focus on remote sensing image and texture image, research on natural image classification is relatively few. On the one hand it is due to the complexity of natural image’s semantics; on the other hand, it is because that most of natural image is colorful, and the color feature is chaotic, making classify images through the contents contained faces greater challenges. Currently the content-based image classification has made some achievements, but most of existing methods is based on single feature of image, then multi-instance learning can be a good solution to this problem. In this paper, based on in-depth study of multi-instance learning and neural networks, we put up a new multi-instance learning method for solving problem of natural image classification.The main contents are as follows:Aiming at natural image classification, we present a multi-instance learning methods based on BP neural network. Through looking the natural images as bags, segmentating regions’ visual features as instances in bags, using K-Means clustering algorithm to sample clustering as input of an improved BP neural network to train a classifier. Experimental results on Corel database show that the algorithm has relatively good performance and higher average accuracy, as well as certain stability. |
