Combining Semi-supervised And Active Learning For Image Retrieval

With the rapid development of multimedia technology and the popularization of the Internet, the number of digital images grows so fast. Image as a rich, intuitive performance, suitable vision multimedia information, its application has been included as national defense, industrial manufacturing, media, health care and public entertainment etc. How quickly and effectively to retrieve the required image from the massive image database increases important. The early 20th century 90, Content-based image retrieval (CBIR) emerged as the times require, it uses color, texture, shape and spatial information to establish the feature vectors of the image to retrieve images. In recent years, the image retrieval system including relevance feedback allow user to participate the determination of the search results, by this way of interaction to create the association of image low-level features and high-level semantics (user evaluation), thereby reducing the semantic gap and improving the accuracy of retrieval.This paper focuses on the research of image low-level feature extraction and relevance feedback, the image low-level feature extraction include color histogram, color correlogram, color-only extraction, Tamura texture extraction, Gabor filter extraction, CEDD-based feature extraction and the FCTH-based feature extraction etc, while the relevance feedback uses the combination of semi-supervised and active learning.Through the experiment, the paper compare with each image low-level feature extraction. In addition, the paper presents a new algorithm which uses the combination of transductive support vector machine (TSVM) semi-supervised and SVMKNN active learning for relevance feedback, experiments show that it greatly improves the precision of each image low-level feature extraction.