Weak Tags Based On Web Image Data Of Spam Image Filtering Method

With the rapid development of computers and the widespread popularity of the Internet, online high-quality digital images are growing. Thus there is an increasing need of new techniques to support more effective image search. Image annotation provides an effective way for people to retrieve the images, however, the auto-annotation technology of image is still not perfect, and prone to error. Therefore, the collaborative image tagging system is now a popular way to obtain large set of labeled images easily by relying on the collaborative effort of a large population of Internet users. In a collaborative image tagging system, people can tag the images according to their social or cultural backgrounds, personal expertise and perception, so the tags may not have exact correspondences with the underlying image semantics, which will result in large number of junk images. Aiming to address the problem discussed above, a junk image filtering algorithm based on weakly-tagged image datasets is proposed in the paper.The junk image filtering algorithm proposed in this paper is based on the image topic network. Our topic network consists of two key components:image topics and their inter-topic contextual relationships. Our paper firstly analyzes the semantic relationships between the weakly tagged images annotation, and extracts the intra-topic semantic context. To achieve more sufficient characterization of the diverse visual properties of the images, both the global visual features and the local visual features are extracted for image content representation and similarity characterization. In our implementations, the global visual features consist of global color histogram and texture features from Gabor filter banks. The local visual features consist of9local color histograms and they are extracted from9simple image partition patterns. The diverse visual similarity contexts between the images are characterized more accurately by combining multiple kernels. The topic network is generated with the image topics and the inter-topic similarity context. In our paper, we introduce some junk image filtering algorithm, including the kernel K-means clustering algorithm used in our paper. The kernel K-means clustering technique is performed to partition the images into multiple clusters and outliers. We use hyperbolic visualization to display the topic network, as well as the junk image filtering results more effectively. After the similarity-based image projection is obtained by using KPCA, Poincar’e disk model is used to map the images on the hyperbolic plane onto a2-D display coordinate. We carry out our experimental studies by using large-scale collections of LabelMe images and Flickr image database.