Local Logo Recognition System for Mobile Devices

In this paper, we propose a novel logo recognition system which can process a very large number of logos locally on mobile devices. The system is not only robust against challenging conditions such as different image scale, rotation, and noisy input, but time efficient and low memory consuming as well. The total computation cost is minimized by using a cascade approach, in which the fast algorithm is kept in the first layer to filter most of the testing cases, while the more expensive but robust one is put on the second layer to investigate only the "confusing" logos. In this paper, we also propose a "background subtraction" method, which considerably improves the second layer in terms of speed, accuracy, and database size. The system has been tested on a dataset of 3000 logos with promising results. The average running-time is just about 1.7 seconds on an average single core mobile device, which is very potential for many mobile applications.