Abstract
As the nexus of visual identity for companies, a trademark serves as the main visual platform for a company’s goods and services. It is important that trademarks be distinguished from one another in identity and/or (visual) similarity to avoid consumer confusion. We conducted a study using two approaches to achieve a deeper understanding of the process behind trademark similarity judgments. First, we used deep neural networks (DNNs) to assess the degree of similarity for trademark image applications. We used the database provided by the Japan Patent Office for the online competition “AI x Trademark: Image Search Competition (Detect similar trademark images)”, containing images that were judged to be similar in appearance in the past by expert officers in real trademark examinations. The dataset consists of a training set size of 2311 images and a validation set size 1542 images. Secondly, we conducted two experiments with human subjects: one behavioral study that was conducted online and one in-person eye-tracking experiment to study how amateur participants visually explore image features and rate the similarity of trademarks. By hypothesizing Japan Patent Office reviewers’ past judgments as “the ground truth” for the DNN classification and comparing the obtained results with experimental results, we investigated the decision-making process to gain more insight into the visual features which are associated with the variability and consistency of professional reviewers’ past decisions on trademark similarity. We report surprisingly low agreement scores (~60%) among DNN/professionals vs. amateur participants (similar to VSS ’22 results), raising psychological and socio-economic issues. By exploring the criteria that participants used in the eye-tracking experiment, we found that shape, color, and orientation are the main determinants when judging overall similarities. The lowest AI similarity scores were mostly associated with images containing various symbols or different geometric figures, such as circles, ellipses, or undefinable designs.