Trademark image retrieval via transformation-invariant deep hashing

Trademark images are usually used to distinguish goods due to their uniqueness, and the amount becomes too huge to search these images accurately and fast. Most existing methods utilize conventional dense features to search visually-similar images, however, the performance and search time are not satisfactory. In this paper, we propose a unified deep hashing framework to learn the binary codes for trademark images, resulting in good performance with less search time. The unified framework integrates two types of deep convolutional networks (i.e., spatial transformer network and recurrent convolutional network) for obtaining transformation-invariant features. These features are discriminative for describing trademark images and robust to different types of transformations. The two-stream networks are followed by the hashing layer. Network parameters are learned by minimizing a sample-weighted loss, which can leverage the hard-searched images. We conduct experiments on two benchmark image sets, i.e., NPU-TM and METU, and verify the effectiveness and efficiency of our proposed approach over state-of-the-art.