Web-based Face Image Retrieval System

With the growing diversity of Internet applications and the rapid growth of Internet information and data, in the actual Internet applications, there are lots of special needs for face image retrieval about celebrities, criminals and popular images. In the traditional text-based image retrieval, the retrieval methods are limited and the retrieval result can be easily deceived by keywords. Moreover, in the general content-based image retrieval, the retrieval precision is susceptible to interference from background image. Both two methods mentioned above are still difficult to meet the needs of the user's retrieval. The retrieval system should combine text messages, image content and knowledge in specific areas to reach a more extensive and more accurate retrieval performance.A new vertical image retrieval framework is proposed to provide more professional and accurate results by considering both text information and image content, as well as the related knowledge in specific areas. Based on this framework, the face image retrieval system is designed and implemented by adding some specific technologies and components, such as automatic filtering of non-human face images and more accurate facial feature extraction and matching algorithm.Through the analysis of the web information and image content, we can filter out those irrelevant pages and those which don't contain any human face image during the data collection process in a retrieval system by using the analysis of the relatively of web pages and face detection algorithm. On the one hand, we can skip those pages during the crawler stage by the analysis in the image context. On the other hand, we use the AdaBoost Cascade algorithm to train cascade classifier to filter out those without any face image.While using the AdaBoost Cascade algorithm to filter out those non-human face images, we can extract face regions from the image. By extracting feature vectors from these face regions, we can avoid those interference caused by the background. Through using the Gabor transform to extract local features and using LBP operator to extract histogram sequence, the similarity matching can be more accurate. Moreover, we proposed a PLGBP algorithm based on spatial pyramid model and the LGBP algorithm, the method improves the retrieval accuracy at the expense of the cost of matching time.In addition, the face image retrieval system also uses Chinese word segmentation , inverted index and relevance feedback techniques. Our system utilize the text semantic information, the image's low-level visual features and the knowledge in face area to obtain an average query accuracy more than 70% and response time within 7s during the query by keyword and example image at the same time from the data set containing 229013 face images parsed from more than one million web pages .