Abstract:
Logo detection is a challenging task with many practical applications in our daily life and intellectual property protection. The two main obstacles here are lack of publ...Show MoreMetadata
Abstract:
Logo detection is a challenging task with many practical applications in our daily life and intellectual property protection. The two main obstacles here are lack of public logo datasets and effective design of logo detection structure. In this paper, we first manually collected and annotated 6,400 images and mix them with FlickrLogo-32 dataset, forming a larger dataset. Secondly, we constructed Faster R-CNN frameworks with several widely used classification models for logo detection. Furthermore, the transfer learning method was introduced in the training process. Finally, clustering was used to guarantee suitable hyper-parameters and more precise anchors of RPN. Experimental results show that the proposed framework outperforms the state of-the-art methods with a noticeable margin.
Date of Conference: 10-13 December 2017
Date Added to IEEE Xplore: 01 March 2018
ISBN Information: