LogoDetectDOCLIB Documentation

Current Version: LogoDetect v1.0

LogoDetectDOCLIB Library is a implementation of the multi-scale logo detection and extraction approach by Zhu and Doermann at ICDAR 2007. It returns the most likely logo candidate on each page whose computed scores are above the specified detection threshold. A higher detection threshold leads to better precision (i.e. higher chance of being really a logo), but may lower the recall. A default detection threshold is used if that parameter is ignored by the user.

The idea is to robustly classifies and precisely localizes logos using a boosting strategy across multiple image scales. At a coarse scale, a trained Fisher classifier performs an initial classification using features from document context and connected components. Each logo candidate region is further classified at successively finer image scales by a cascade of simple classifiers, which allows false alarms to be quickly discarded and the detected region to be refined.

We assume that the logo in each document appears on the top one third of the document. If more than one logos are present on a document, the one with highest computed score is selected.


LogoDetectDOCLIB Library is an add-on of DOCLIB. DOCLIB is being developed under contract by a collaboration between:
The Laboratory for Language and Media Processing
Unviersity of Maryland, College Park
and
Booz | Allen | Hamilton

All Rights Reserved, 2003-2007