Aggregation of reversal invariant features from edge images for large-scale trademark retrieval

Abstract

Trademark retrieval is of wide application in many fields and has motivated lots of researchers to study on it. In this paper, we proposed a pipeline of trademark retrieval method including three steps: extraction edge, split image into significant components and extract enhanced SIFT features. Our work firstly proposed extracting features from trademarks' edge images, which can obviously boost the retrieval performance. Moreover, we proposed a simple but effective algorithm which can split trademark images into several significant components, this step is important so as to tackle the local similarity problem in trademark retrieval task. Then we extract reversal invariant SIFT features from these components images with Fisher Vector (FV) aggregation step. We conducted our experiments on a large-scale and challenging trademark dataset METU v2, results show that our proposed method based on hand-craft features even outperform the methods based on deep learning methods (CNN mainly) and achieve the state-of-the-art performance.