Rapid Estimation for Logo Detection on Mobile Devices

Abstract

Along with the technological development, running complicated program on mobile phone become possible. Mobile device allows to easily capture pictures and do corresponding processing. Thus, this paper puts forward a rapid estimation method which is aimed at detecting large-scale logos in the natural environment by using mobile device. For this purpose, feature detectors and objectness measure are applied to rapid estimation method. By using the feature extraction results to evaluate the objectness results, the higher the evaluation is, the more possible the objectness result is a logo region. This paper take 1000 pictures in the natural environment for measurement, the experiment results show the effectiveness of the method. In the last few years, logo recognition in the reality environment has drawn more and more attention. Currently the methods on the logo detection are mainly training a classifier with the global feature or local feature which is extracted from certain logo, and finally get the model and classification equation. When using these methods to deal with a large number of logos in the natural environment, it needs lots of training images that will bring a huge amount of computation. Thus, this method cannot meet the demand for real-time logo detection at the present stage. Now with the development of intelligent handset technology, mobile phone has a good inner set camera and processor, such as human face detection on mobile phone is already a mature technology. Therefore, using mobile device to get to the logo images and do processing work to complete the function of logo detection has become possible. This paper introduce a fast estimation method, which can estimate the possible logo region in an image. This method is divided into two parts. The first part consists of feature extraction and object detection. Second part is a screening process, the results of feature extraction are used to screening the target windows of the object detection results. In this paper, FlickrLogos-32 dataset and 1000 pictures taken in the natural environment are used for testing, the experimental results indicate that the method is effective.