Research On Annotation Technology Of Face Images In Network News

With the rapid development of the multimedia technology and computer technology, multimedia data has become an important information resource. People can get access to large amounts of multimedia information, such as video, images, various types of news through the Internet everyday. As a kind of multimedia data, network news is related to the behaviors and activities of public figures and the behavior of public figures are closely related to the current important events. Therefore, if public figures in the news photos can be easily retrieved, people can quickly understand the dynamics of the events. If the characters in the news images can be annotated and a database can be generated, people can easily and quickly retrieve the characters. Annotating the characters manually is effective, but the rapid development of multimedia and network technology makes the image library very large and annotation a heavy work.Images and text in network news are interrelated. In this paper, the characteristic of network news that images and text complement each other is used to produce a database of face images that labeled by names. A lot of news data can be got from the network. Names of the public figures can be extracted from the text of the news data and face images can be got from news images related to the text by face detection. A lot of face-name pairs with noise can be got ultimately and the right corresponding relationship of faces and names can be got by removing the noise with machine learning algorithms. Characters that appear in the news images are automatically labeled with their names. Image retrieval is converted to text retrieval to solve characters retrieval better.A face detection algorithm based on Adaboost was used to get a lot of face images. The quality of face images was improved through median filter and histogram equalization and face images were standardized normalized by eye detection. Finally a database of normalized face images was got.AP clustering algorithm is improved considering the characteristic of network news that images and text complement each other. A face annotation method based on improved AP clustering algorithm was proposed to find the correspondence between faces and associated names. A cluster merging process was proposed to identify the same person with different names. The improved AP clustering algorithm can make full use of the complementary feature of images and captions in the network news to get more accurate correspondence between the faces and names. A database of face images accurately labeled by names which supports content-based character retrieval and text-based character retrieval can be got.