Research On Key Techniques Of Content-Based Image Retrieval And Content-Based Image Filtering

Building up content-based image descriptions, retrieval and filtering on the Internet is a crucial issue because of the recent explosion in the amount of online images. By using the technologies of image processing, pattern recognition, computer vision and database, this dissertation studies some key problems in the field of image retrieval and image filtering. In order to improve the performance of CBIR system, the research work concerning the image retrieval field of this dissertation focuses on relevance feedback method and semantic model oriented to image retrieval system with relevance feedback. In the following part of the dissertation, the concept, architecture and model of content-based image filtering has been discussed, and adult image filtering methods have been studied in order to reject offensive embedded images on web pages. The contributions of the dissertation are as follows:(1) This dissertation presents a relevance feedback method for image retrieval —Rich Get Richer (RGR), which is based on Bayesian inference. It takes the time sequence characteristic of relevance feedback into account, and modifies the effect of relevance feedback to image retrieval system. So, this system has the ability to be adaptive, which makes the retrieval result consistent with the user's subjectivity. The experimental results have shown that the proposed approach can capture the user's information need more precisely and quickly.(2) An image internal semantic model for the relevance feedback image retrieval system is proposed. It extracts the semantic information by analyzing relevance feedback image retrieval results without troubling image annotation and difficult image segmentation (Image segmentation is an open problem). In order to get the semantic correlation of images, two learning methods based-on mutual information and image association factor, are proposed.(3) This dissertation presents two semantic clustering methods. One is based-on mutual information learning method, and the other is based-on association rule hypergraph partitioning algorithm. Different from the clustering method based-on visual feature, it analyzes the historic information of user retrievals and creates a semantic related image translation database. This provides a new way for image database management.(4) For content-based image filtering, the dissertation presents a multi-layer filtering method, which creates an efficient color skin detection model according to the remarkable characteristic of adult images, and also uses support vector machine and K-nearest neighbor methods to reject offensive images. Experiments have shown that the method is efficient for adult image detection, but many normal images are also recognized as adult image. Therefore, the multi-feature filtering method is designed to resolve the problem. Experiments have shown that the multi-feature filtering method is very effective.