Automatic collecting representative logo images from the Internet

With the explosive growth of commercial logos, high quality logo images are needed for training logo detection or recognition systems, especially for famous logos or new commercial brands. This paper focuses on automatic collecting representative logo images from the Internet without any human labeling or seed images. We propose multiple dictionary invariant sparse coding to solve this problem. This work can automatically provide prototypes, representative images, or weak labeled training images for logo detection, logo recognition, trademark infringement detection, brand protection, and ad-targeting. The experiment results show that our method increases the mean average precision for 25 types of logos to 80.07% whereas the original search engine results only have 32% representative logo images. The top images collected by our method are accurate and reliable enough for practical applications in the future.