This project set out to build a content based image retrieval (CBIR) framework for a collection of handwritten documents with the goal of being able to find specific images of words within a large collection. Background research showed that a system of this nature has never been built before. A framework was designed in order to allow for the implementation of systems of this type for any collection of handwritten documents.

Evaluation showed that each component of the framework contributes to returning results and that the components act together to produce good results. An attempt was made to try to identify optimal values for variables used in the system and, while some insight was gained, further investigation would need to be done in order to gain a more complete picture.

Time constraints meant that not all possibilities for the different components of the system could be implemented but those which were implemented showed that success using this type of system could be achieved though scalability becomes an issue and the system tends to operate slower and less accurately on larger collections.

The system is experimental and not yet ready for use by end users but results up to this point are promising.

In answering the research question - can image-based searching be done accurately and efficiently? - it was found that when a good search key is selected, then relevant matches can be found thereby suggesting that it is possible to do image-based searching accurately when a good search key is selected. Similarly, it was shown that the system performs well with matches taking approximately 1 second on a collection size of around 3000 images and 16 seconds on a collection size exceeding 14000 images. These times are reported without any speed optimisations having taken place and it is suspected that speed optimisations could significantly improve the speed of the system.


The BOLD Translator

by Kyle Williams







Design and Implementation






Future Work