Logo Recognition System Based On Saliency Detection

With the development of the Internet, image data increases dramatically. How to analyze or organize the huge image data and extract useful information from it has become a hot topic in today’s study. In order to solve this problem, content-based image retrieval emerged. And logo recognition is a very important part of it. The development of social network and economic globalization gives the logo recognition great commercial value and let logo detection and recognition have more practical significance. For this reason, our article has proposed and implemented a logo recognition system based on saliency detection.Different image type has different logo recognition method. The image type can be divided into document image and real-life image. Our logo recognition system focuses on real-life image and we need to solve the following two major problems. The first one is how to position logos accurately while logos in real-life images is usually small and incomplete with various morphological transformation. The second one is how to describe the various logo appearance of the same logo in real-life image to match the correct picture.Logo recognition system was divided into logo detection module and logo recognition module in order to solve the two problems above. The purpose of logo detection module is logo positioning and the purpose of logo recognition is solving the matching problem. In the logo detection module, we first use a saliency detection method to generate the saliency map of source picture. Then we propose a logo location method based on the saliency point distribution of the salience map to locate the axes of logo. In the logo recognition module, we use SIFT feature and bag-of-word feature to describe the characteristics of logo area. Then an inverted index has been used to organize and retrieve the image dataset. Finally we get the recognition result based on the retrieval of test image.According to the experimental results, our logo location method can locate the image logo well and the system can match the real-life image correctly. We have reached the expected purpose.