Mobile product recognition for information retrieval

Abstract

In this final year project, theoretical background of Bag-Of-Words were introduced, various aspects and performance analysis of product recognition system was explored.

Two databases were constructed for this project, one is mobile database, and the other one is internet database. In mobile database, 96 product logos were captured from different views using different mobile phones, whereby in internet database, 15 product logos were downloaded from Google and Flickr. Both databases were used for system performance analysis.

Keypoint and dense sampling these two different feature extraction approaches of the system were used to compare the performance in terms of accuracy and time.

Vocabulary Tree (VT) approach was used. Geometric verification step was added after VT to improve the accuracy of the recognition system. GV using Harris corner detection and David Lowe’s SIFT keypoint detection approaches were compared.

To further understand the performance of database, low accuracy database was studied and matching points for a few examples are shown using both approaches.

In summary, the product recognition system was explored and analyzed. A comprehensive and representative mobile database was built. Keypoint based BoW framework, combined with geometric verification which using SIFT keypoint detection was shown to provide the best performance both in accuracy and speed; the product recognition system is practical and promising in the future application of information retrieval.