Identification Of Important Person From News Image Collections Based On Cross-modality

As the network news data grows explosively, how to quickly and efficiently retrieve news information from the mass of data becomes an urgent problem to be solved. Most of network news is related to some persons(especially important persons, such as foreign dignitaries) and their activities, and the graphic information and text information in network news can be used to make an association analysis, especially make an association analysis between person face image and person name in network news, which can lead to an automantic analysis and information mining about news events. Making an association analysis between person face image and person name is becoming a leading international research focus in recent years.The correspondence between face images and names in news has often many-to-many relationships, so this makes it extremely difficult to realize automatic news person identification from the multiple correspondence relationships. In this paper, a special study on automatically identifying persons in network news is performed. It is based on a cross modality information fusion method using image and text information. The main research work includes:1. For the problem of giving an important person name to find the corresponding person face image in news, a naming algorithm based on cross-modality fusion is studied. In the problem, traditional identification method is based on text label, while the text-based identification methods existed label ambiguity and image-based identification methods existed visual diversity problem because of many different factors(such as expression, illumination, pose and occlusion). As the order of names appeared in news caption makeing great contribution to correctly identification of news person, so this thesis proposed an algorithm that fuses the position(order) of names in news caption, the visual information and the similarity of face images for face naming. We performed experiments on the data set which consists of approximated half a million news images from Yahoo news. Compared with the traditional Berg’s algorithm, the recall and precision of new algorithm respectively increase 8.9 percent and 43.5 percent.2. For the problem of giving a face image to find the corresponding name in news, a face annotation algorithm based on improved Max-ED and Imax-ED is studied. In the problem, manually labeling the positive and negative bags of news images set was needed and that was time-consuming and labor-intensive. In this paper, an improved ED and Iter-ED algorithms of face annotation in video news were introduced into person annotation in news images, and on the basis of reducing the influence of false-positive bag for annotation, an improved assignment method of positive/negative bag was proposed. The experimental results based on FAN-Large data show that the accuracy of automatic identification of the proposed algorithm increased 42.4 percent.3. For the problem of constructing a names-faces correspondence relationship, a personal identification algorithm based on a name semantic network building on network global information is studied. Considering that people that appeared together in the same caption are high probability emerging together in the same news images associated with the caption, in order to improve the performance of automatic personal identification algorithm, the thesis proposed the algorithm of building a name semantic network based on network global information, and personal identification algorithm based on a name semantic network. The experiments were performed on the data set which consists of approximate half a million news images from Yahoo news. Compared with Ozkan’s algorithm, our algorithm can resolved the problem that intra-class distance among face images is reduced by a number of factors, thereby the precision and recall of our algorithm respectively increased 22.1 percent and 24 percent.