Robust Logo Detection Across Large Style Variations

Zhiyuan Zhao, Qingjie Liu

Abstract: Style variation of logo refers to changes in the logo’s visual characteristics during the evolution of the logo, which is a common yet easily overlooked phenomenon. However, conventional logo detection methods suffer from severe performance degradation once the visual characteristics of the logo change, because they fail to establish a relation between different styles due to their lack-of-interaction learning procedure. In this paper, we attend to address this detection failure by learning a transferable and flexible cross-style relation under the meta-learning policy. Our proposed method contains one more sibling branch except for the vanilla Faster-RCNN pipeline, which creates a pair-wise comparing environment. Meanwhile, the classification head of the detector is remodeled into a matching module which meta-learns how to classify regions through pair-wise matching. This pair-wise matching mechanism gives matching module the ability to establish deep transferable relations across styles. Additionally, two logo detection datasets are proposed to support research on logo detection across style variations. Experiments revealed the superior performance of our proposed method.