Abstract
Automatic logo detection is a key function for many applications. Most existing logo detection methods are based on strong object level bounding box annotations, annotated either manually or automatically in synthesized images. However, in real applications, there are thousands of different types of logos with clutter background and varying sizes in images, that makes supervised detection methods less applicable because it is a labor-intensive task to do bounding box annotations. On the other hand, some weakly supervised learning-based methods have been proposed, their experimental results are not promising. In this paper, we propose weakly supervised methods for logo detection based on a dual attention dilated residual network (DRN) containing spatial and channel attention mechanisms using image-level labels instead of bounding box annotation for training data. The incorporated spatial attention module computes attention weights which are useful in predicting the spatial location in an image. Channel wise feature maps are used to emphasize the dependency of the channel in the global feature map and help to classify the logos into different categories. The use of attention-based mechanisms with DRN improves classification accuracy (around 4%) and considerably increases localization accuracy by more than 4% over a regular dilated residual network architecture.
