Research On The Techniques Of Semantic-based Image Retrieval

Semantic-based image retrieval, as a more reasonable image retrieval technique, searches images according to user's understanding, and it has become the major tendency in content-based image retrieval field. Through the introduction of image semantic model, image semantic extraction and description, this paper discusses some existing problems and certain solutions of semantic-based image retrieval. Focusing on the image semantic classification and semantic-based retrieval, this paper conducts a study of design and implement of semantic-based image retrieval system. And the main contributions are as follows:Based on the initial classification of the image database, various visual features of images in training database are extracted, and the application of SVM (Support Vector Machine) in the hierarchical semantic classification is presented. Then certain SVM parameters and normalized feature vectors are used to train the SVM classifier, of which the influence to the semantic classification is analyzed based on recall and precision. The result of semantic classification is a semantic determinant tree, certain semantic rules are gained for every semantic classifier, and every semantic concept in the database can be determined by some semantic rules. The results show that appropriate visual feature and SVM parameter are helpful for the correct semantic classification of image database.A prototype of semantic-based image retrieval system is realized, it can receive the query mode by both keyword and example image. In the keyword query model, relevance feedback algorithm based on SVM incremental learning is brought forward. Considering both the relevant results and the irrelevant ones, SVM classifier is re-trained, the semantic rules are modified, and the retrieval performance is improved. In the example image query model, aimed at every level of classifier of the semantic determinant tree, corresponding prominent features of the example image are extracted, and its semantic sort can be estimated. The experimental results show that the retrieval system can fit the user's semantic and visual demand at a certain degree by the integrated application of both query models and the relevance feedback.