Interactive mobile media search engine

Abstract

The proliferation of mobile devices with high-resolution built-in camera has led to a new form of information searching: the input of search engine is no longer restricted to text only, but extended into the form of image and voice as well. In particular, mobile visual search has been widely studied by researchers to develop the new kind of search engine that is adaptable to image-based input. Different approaches have been proposed to optimize the process of feature matching and geometric verification (GV) of images so that the retrieval result is accurate and efficient.

The objective of this thesis is to optimize the performance of mobile visual search system in terms of classification accuracy and time performance. The goal was achieved through analyzing the problem of the current GV, as well as, investigating the parameters or thresholds used in pairwise matching and RANSAC estimation algorithm. Different variations of robust estimation of geometric transformation were also explored as the possible alternatives other than RANSAC, such as PROSAC, MLESAC and SCRAMSAC.

We have addressed the problem in the implementation of mobile visual search as well as the necessity to speed up the process of GV while maintaining a considerable improvement on the classification accuracy.