Siamese Neural Network Optimization Using Distance Metrics for Trademark Image Similarity Detection

Abstract

: This study focuses on optimizing Siamese Neural Networks using various distance metrics to enhance trademark image similarity detection. Traditional Euclidean methods often fail to detect subtle visual differences, leading to less accurate outcomes. This research incorporates Chi-Squared and Manhattan metrics into the network, in addition to the conventional Euclidean metric. Using 255 trademark images across five classes, triplet samples were created for training and evaluation. The models, which utilized the CNN Xception architecture and a triplet loss function, were trained separately with each distance metric. Performance was assessed comprehensively via multiple metrics, including accuracy, precision, recall, and F1-Score, to ensure robust evaluation. The results indicated that the Chi-Squared metric significantly outperformed the others, achieving an impressive accuracy of 0.96. In comparison, the Manhattan and Euclidean metrics achieved accuracies of 0.74 and 0.92, respectively. Notably, the Chi-Squared metric improved accuracy by approximately 4.35% compared to the Euclidean metric. These findings underscore the critical importance of selecting suitable distance metrics for image similarity tasks, as the choice of metric can substantially impact performance. The Chi-Squared metric was particularly effective due to its sensitivity to variations in features such as color and texture, which are often pivotal in trademark images. This research demonstrates the substantial benefits of incorporating advanced distance metrics into deep learning models for trademark similarity detection, providing a more accurate and reliable approach. By highlighting the effectiveness of the Chi-Squared metric, this study paves the way for future research to further refine these metrics and explore their applications in various domains requiring precise image similarity analysis. Future studies may also consider integrating additional metrics or hybrid approaches to further enhance performance and applicability in diverse contexts.