Abstract
Mobile visual search (MVS) has become a hot topic in both academia and business circles. A typical MVS system generally uses the camera phone to initiate search queries, following a client-server architecture. Transmission overload turns out to be a bottleneck due to the bandwidth constrained wireless link. An effective way to reduce transmission overload is exploiting progressive transmission strategy, which probably obtains the correct match via a few more important image descriptors delivered. However, the acceleration effects of existing methods are often limited, because they only decrease the amount of descriptors delivered, and important descriptors with high dimension still need being entirely delivered. To compress the descriptors and further reduce network latency, many methods have been offered, such as PCA-SIFT, SURF, CHoG. One of the problems is that they cannot restore the original descriptor and generally the more compressive the descriptor is, the lower the precision is. To conquer the above problems, we propose a progressive transmission strategy based on multi-level wavelet decomposition and reconstruction theorem, which can both compress the descriptor and restore the original data at the same time. Extensive experiments have been done on the public Stanford MVS database, demonstrating that our proposed progressive transmission strategy outperforms other strategies based on SIFT and SURF descriptors directly, when delivering the same amount of data.
Yegang Du
Assistant Professor
My research interests include intelligent system, HCI, AIoT, and pervasive computing.