Content based image retrieval with LIRe
- 1. CBIR with LIRe Mathias Lux Klagenfurt University Austria This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0
- 2. Lucene Image Retrieval • Open source • Java • Based on Lucene – text retrieval engine
- 3. History • Caliph & Emir – first release in July 2004 – developed since ~ midth of 2003 • Lire as spin-off product – first release in 2006 – first use case was „barrique“ – Just MPEG-7 Barrique is a project of Roman Kern: http://sourceforge.net/projects/barrique/
- 4. Image Retrieval Features • Color Histograms • MPEG-7 features – scalable color, color layout and edge histogram • Tamura features – coarseness, contrast and directionality • Joint histograms CEDD, FCTH and JCD • Auto color correlation • JPEG coefficients histogram • Local features – SIFT , SURF, MSER • ... and more to come
- 5. Indexing and Search • Fast linear search • BoVW search based on Lucene • Parallel k-means & visual word indexing for BoVW • Fastmap for spatial indexing • Metric spaces approach for fast search in global features
- 6. Lire: Some Examples • Utilize Lucene – put an index into RAM – distribute indexes over the network • Rapid prototyping – Basic auto-tagging < 100 loc • Test different features & approaches – eg. mediamid & CEDD
- 7. Where has LIRe been used? • ImageCLEF – different groups – mentioned on the website • TRECVid – 2010 ranked 3rd in known item search • Medical domain – image retrieval & video summarization • Apparel search applications
- 8. Future Directions … • Unified way to change & test metrics • Easier benchmarking • Further support and maintenance
- 9. Thanks … • Want to download & try LIRe: • http://www.semanticmetadata.net