Research On Key Techniques Of User Oriented Web Image Retrieval

With the rapid development of computer technology and network technology, the volumes of Internet information are increasing continuously. How to find user interested images quickly and accurately from huge web image resources has become a great challenge. Currently prevalent approaches to web image retrieval fall into two main categories: text-based image retrieval (TBIR) and content-based image retrieval (CBIR).TBIR is the main method that current commercial image retrival engines used. However, its confronted problem is that it only used textual information around web images, and does not include image content information. While CBIR confronted problem is semantic gap, i.e. low-level visual features such as color, texture, shape etc. cannot describe high-level concepts effectively.The web images in the web context have multi-modal characteristics obviously. It is still research hotspots in image retrieval field that how to fully utilize the above multi-modal information for satisfying users'retrieval demands.To address the above issues, this dissertation investigates the key techniques on web image retrieval from several perspectives, including image annotation based on fusion of textual and visual information, semantic gaps between low-level visual features and high-level concepts, image retrieval based on user interest model, relevance feedback technique etc.Firstly a user-oriented image retrieval framework is proposed.The system can support multi-modal information query. It uses SVM to build and update semantics in combination with relevance feedback. The twice distance of images is computed to reinforce the semantic discrimination between images and make the images retrieved more similar to query image from the semantic point.A new 20 color none uniform quantization method based on HSV space is propesd. Compared with previous algorithm, it reduces feature dimensions, and improves retrieval precision. Then we train SVM classifier using visual features such as color, texture, and shape. Experimental results show that this approach can improve precision and recall effectively.An automatic annotation algorithm based on text and visual features fusion for web images is proposed in order to improve the reliability of web images annotation. We first extract the image context information, and then include image title, page caption, image URL, image ALT into feature assembly. The candidate textual keywords are extracted using WordNet semantic dictionary. We then use Mean-Shift image segmentation algorithm, and extract visually salient areas with saliency map. We filter the annotation words by relativity between visual features and annotation words . We cluster image semantics classes using PLSA method. Finally, we build relations between textual keywords and visual feature classes by semantic network to realize automatic annotation.We introduce user interest model into image retrieval field. Our user interest model combines explicit tracking and implicit tracking to improve user's interest information and provide personalized information retrieval service. Experimental results show that this approach can improve precision and recall effectively.Finally, a user-oriented image retrieval system is finished. The conclusions and the research prospects of the thesis are presented.