Mobile Retriever - Finding Document with a Snapshot

In this paper we describe a camera based document image retrieval system which is targeted toward camera phones. Our goal is to enable the device to identify which of a known set of documents it is “looking at”. This paper provides two key contributions 1) a local context descriptor that effectively rules out irrelevant documents using only a small patch of the document image and 2) a layout verification approach that boosts the accuracy of retrieval even under severe degradation such as warping or crinkling. We have implemented the mobile retriever client on an iMate Jamin camera phone and tested it with a document database of 12742 pages. Experiments show that our verification approach clearly separates successful retrievals from unsuccessful retrievals.