One in a billion: finding matching images in very large corpora
summary
An open-source reverse-image-search system called image-match claims a complete free solution that encodes perceptual hashes, uses Python and numpy, and indexes vectors into Elasticsearch. It is built to sustain 10k inserts/s and to return nearest neighbors by breaking signatures into integer indexes; empirical desktop tests show logarithmic lookup and authors extrapolate 1-2s lookup at 1B images though a billion-image cluster is not demonstrated. The algorithm follows an early 2000s perceptual-hash paper and the code is available on GitHub.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:04Perceptual image similarity
- 2:47Image signatures from grids
- 5:21Measuring image distance
- 6:45Efficient large database search
- 10:13Method performance and robustness
- 12:00Related projects and open-source
- 13:24Acknowledgments
- 14:30Q&A
related talks
26mSimilarity search: Take your search engine to the next level
33mArchitecting Scalable Multi-Modal Video Search
37mSimilarity Search for Product Matching @ Semantics3
01h 28mBuilding a Semantic Search Engine
29mUnleashing the Power of Vector Search and Semantic Caching in Java
11mSimilarity Search Application utilizing Vector types on vCore-based Azure Cosmos DB for MongoDB
55mVector Similarity Search in Spring with Redis Stack
22mBuilding a Multimodal Search Engine