Mobile landmark recognition for information retrieval

Abstract

The final year project report aims to cover the theoretical background of the existing recognition techniques, different components for recognition process and performance analysis of different combination of feature detection technique to use in Mobile Landmark Recognition application.

Content-based image retrieval has been researched upon with techniques ranging from image feature extraction, representation, indexing. Bag-of-Words framework is used in the project where each image is modeled as a collection of features plotted over a histogram. A database with 50 categories containing a total of 4000+ images is used to train and test the system. This database was built with image recognition in mind so as to ensure that it is able to simulate the real scenario where users capture image using their smartphones. Two feature detection methods and two different resolutions were used. Two feature detection methods were dense sampling based and keypoint detection based approach. 640 x 480 and 320 x 240 resolutions were used in the project. Hierarchical k-means approach is used in clustering and scalable vocabulary tree is used in machine learning.

In summary, it is proven that using dense sampling based approach performed better than keypoint detection based approach. Even reducing the resolution still gives the best result for dense sampling.