Design And Implementation Of Image Retrieval System MOVER

Content-based Image Retrieval Engine plays an important role in general image retrieval engine, copyright intrusion detection and video object tracking. It also counts in the field of security such as illegal image filtering. This paper explores deeply the problem of content-based image retrieval and build a system called MOVER (Mobile Visual Engine of Retrieval) which supports mobile client retrieval and utilizes SIFT local feature, LSH high dimensional indexing mechanism and two-view segmentation technique. This system has a test-automation framework and a lot of improving methods are conceived and implemented in the system.First of all, distributed architecture is designed to improve performance and support more images at the online retrieval stage. Following methods are researched and used by this paper: 1. Image database pre-classification is done based on SVM or bag of words techniques. 2. Stable points are filtered out using the combination of simulating user-taken images and two-view segmentation to optimize high dimensional index. 3. PCA analysis method is used to decrease the size of SIFT. 4. Multi-feature combination method is used to compliment the drawbacks of a single feature. 5. Min-hashing technique is used to support large-scale image retrieval. 6. Machine learning is used to improve the image relativity scoring method. 7. Test-automation method is deployed to locate the drawbacks of the system swiftly.This paper has done a lot of experiments to test the system's precision-recall capability. This system has a good precision over 80% while the recall is above 50%. The average retrieval time for a middle or large image is about eight seconds which is acceptable. LSH optimization method is a key test point and over 13.83% index memory space could be save while the accuracy is reduced only by less than 1.5%.