Similar Trademark Image Retrieval Based on Convolutional Neural Network and Constraint Theory

Trademarks are intellectual and industrial properties developed under the commodity economy, representing reputation, quality and reliability of firms. Therefore, in order to prevent the registration of new trademarks from having a high-degree similarity with registered ones, we propose a new trademark retrieval method. Based on the fact that the shape and color of a trademark are varied, our proposed method combines a metric convolutional neural network (CNN) and conventional hand-crafted features to describe the trademark images. More specifically, we first train the CNN based on Siamese and Triplet structures, and then extract the hand-crafted features from convolutional feature maps. For this research, we utilize a challenging trademark dataset that contains 7139 various color or gray images. Besides, extensive experiments on our dataset and the METU public dataset demonstrate the effectiveness of our method in trademark retrieval and achieve the state-of-the-art performance compared to traditional countermeasures.