Abstract
Global trademark applications have risen by approx. 46% from 2017 to 2020, according to the World Intellectual Property Organization. As the primary source of visibility for companies, a trademark symbolizes the company’s identity to distinguish itself. Therefore, similarity and confusion are always an issue that raises intellectual property issues. Here, we evaluate the consistency and appropriateness of professional trademark examiners’ judgment on image similarity. First, we use a machine learning approach. We utilize the database provided by the Japan Patent Office for the open online competition “AI x Trademark: Image Search Competition (Detect similar trademark images)” (scheduled on 11/26/2021 to 1/31/2022) to train the DNN to mimic the real judgment records. The database contains images that were judged to be similar in appearance by an officer in real trademark examinations. Using officers’ past judgments as the “ground truth”, we use 2311 pairs of identical trademark images that are judged as similar into the DNN and then test them on 1542 pairs to build a classifier system for trademark similarity judgment. Using the resulting classification, we assess the variability and consistency of professional reviewers’ past decisions. There is consistency, as well as some “oddballs” (i.e., cases with discrepancy), providing hints towards an optimized, hybrid AI decision aid. Next, our psychophysical tests with naïve observers further reveal differences between professional and naïve judgments on image similarity. Thus, we provide an objective machine framework based on which we can evaluate both professional and amateur reviewers’ performances.