Research On Region-Based Image Retrieval

The traditional text-based image retrieval technology mainly depends on manual labeling, which has the disadvantage of large amount of work, time-consuming and subjectivity. Content-based image retrieval technology appeared in nineties can make up for the limitation of it to some extent. However, because there is obvious difference between the image feature and subjective apperception, and single feature can not express the image sufficiently, the content-based image retrieval technology is hard to satisfy the retrieval demands of users. Therefore, Multi-feature-based image retrieval technique and region-based image retrieval technique have been put forward recently.This paper focuses on the research of region-based image retrieval technique and mainly discusses the algorithm of image salient point detection. The color and texture feature around the points are extracted for image retrieval. The main contents are as follows:At first, a new color image salient point detector is proposed through properly improving the original Harris interest point detector. Compared with the Barnard, Forstner, SUSAN interest point detector and the edge detectors, this improving algorithm has a better distribution effect.Secondly, applied the wavelet transform to image salient point detection, a new color salient point detector based on wavelet transform and Barnard detector has been proposed. This method combines the merit of wavelet transform and the Barnard detector. It overcomes the shortcoming that salient points are easily assembled in some small regions when comparing with the edge detectors and interest point detectors. The salient points extracted by this method are extremely effective on expressing different image regions.Finally, the two new salient point detectors are used for image retrieval and a new region-based image retrieval method is proposed. Color differential invariants, color moment, Tamura texture and co-occurrence matrix are extracted in the region near the salient points. The retrieval results are given by comparing the query image features and the image features in database. Experiments show that the points extracted by the two new detectors are more dispersive than others. When applying them to image retrieval, higher accuracy is achieving. And the salient points extracted based on wavelet transform and Barnard detector can get much better results than classical methods when used in image retrieval.