| Content-based image retrival systems have been widely used with the developmentof multimedia and Internet technology. However, how to automatically detect and targetimages in an image database under the changes of light, geometry and still meeting theneeds of the user’s visual image, is a challenging research topic. This thesis makes asurvey of image retrival system based on local invariant features.Firstly, content-based image retrieval system has been studied, and a conclusion oftypical architecture of the system has been made. Besides,a brief overview of thefeedback that are involved in the system and user interface technologies have beenmade.Further more,the criterias for evaluating the performance of CBIR have been listed,thus contributing to further study in image retrieval sytems.Secondly, the research of features extraction technology based on local invariantfeatures has been deeply made. A robust SUSAN corner points detector has beenintroduced and an improved L-SUSAN algorithm proposed for reducing the number ofcorner points for efficiencily image matching later; Another elegant blob detector calledSIFT has been introduced in detial and an improved SIFT descriptor C-SIFT,based onconcentric round template of simplified description, has been invented for its too muchfeatures points extraction and high degrees dimension. Experiments proved that theL-SUSAN algorithm both guaranting the image matches and greatly reducing the numberof corner feature, significantly improved the efficiency of matching.Thirdly, the feature matching technology has been also conductedin-depth,especially focusing on an approach based on nearest-neighbor and next-nearestneighbor distance ratio of similarity measure method based on improved BBF KD treenearest neighbor search strategy. After that, experiments has been made to compare SIFTand C-SIFT descriptors. Experiments have shown that improved C-SIFT descriptorshares the same ability which is invariant to scales, rotation and illumination changes butbrings a greatly improving the efficiency of algorithms computing and image matching.Finally, an image retrival system based on local invariant features has been designedand implemented according to the the above research. |
