Image recognition and registration are among the most commonly used technologies in mobile multimedia applications. With the introduction of robust local image features, both the accuracy and robustness of image recognition and registration have been improved greatly. However, these features are expensive to extract, and are too slow for real-time
applications on a mobile device. For example, the popular MSER/SIFT algorithm extracts feature locations
based on maximally stable extremal regions (MSER), and calculate a rich feature descriptor provided by SIFT[11]. However, the computation is too challenging for a mobile device. Furthemore, being characterization of local signatures, the feature descriptors can be locally ambiguous. For certain image regions, this problem becomes more prominent.