Trademark Logo Infringement Detection: A Threshold Determination Approach Using Bayesian Optimization and Siamese Neural Networks

Trademark infringement detection is essential to prevent consumer confusion, particularly in the digital era where visually similar logos are widespread. This study proposes a similarity threshold determination framework using Siamese Neural Networks (SNN) combined with Bayesian Optimization to improve the accuracy of trademark similarity assessment. Logo images were collected from the Indonesian intellectual property database, preprocessed into a uniform format, and trained using a triplet loss approach. Bayesian Optimization was applied to determine the optimal similarity threshold, minimizing false positive and false negative classifications. The proposed model achieved an accuracy of 92.23%, with precision of 93.34%, recall of 91.44%, and an F1-score of 92.39%. The optimal threshold (0.200149) effectively balanced sensitivity and specificity, resulting in low misclassification rates. These findings demonstrate that integrating SNN with Bayesian Optimization provides a robust and legally relevant framework for trademark infringement detection, offering practical implications for strengthening intellectual property protection.