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

Suyahman, Suyahman and Sunardi, Sunardi and Murinto, Murinto and Nur Khusna, Arfiani (2025) Siamese Neural Network Optimization Using Distance Metrics for Trademark Image Similarity Detection. [Artikel Dosen]

Text Siamese Neural Network Optimization Using Distance Metrics.pdf - Published Version Download (2MB)

Abstract

Trademark image similarity detection plays a crucial role in protecting intellectual property. Traditional methods, particularly those relying on Euclidean distance, often fail to capture subtle visual differences, leading to less accurate results. This study addresses this issue by optimizing a Siamese Neural Network (SNN) with improved distance metrics. Specifically, Chi-Squared and Manhattan
distance methods are explored alongside the standard Euclidean approach to enhance trademark similarity detection. The objective is to develop a more precise and reliable system for trademark analysis, essential for effective intellectual property enforcement. The research
utilizes a dataset of 255 trademark images across five classes, each with variations in color, texture, and design. To train and evaluate the model, 2000 triplet samples—comprising an anchor image, a similar (positive) image, and a dissimilar (negative) image—were generated, with 1600 pairs used for training and 400 for validation. The SNN model was built using the Xception CNN architecture
and trained with a triplet loss function to distinguish between similar and dissimilar images. Performance was assessed using accuracy, precision, recall, and F1-score. Results demonstrated that the Chi-Squared distance metric outperformed the others, achieving an accuracy of 0.96, compared to 0.92 for Euclidean and 0.74 for Manhattan. The Chi-Squared metric proved particularly effective in
capturing differences in color and texture, improving accuracy by 0.0435 over Euclidean. These findings highlight the significance of selecting appropriate distance metrics for image similarity tasks, as they directly impact performance. This study advances traditional trademark similarity detection by integrating optimized distance measures, making automated trademark protection more reliable. Future research may explore hybrid metrics or novel approaches to further improve accuracy across diverse trademark datasets, strengthening legal and business efforts in safeguarding intellectual property.

Item Type:	Artikel Dosen
Subjects:	T Technology > T Technology (General)
Divisi / Prodi:	Faculty of Industrial Technology (Fakultas Teknologi Industri) > S1-Informatics Engineering (S1-Teknik Informatika)
Depositing User:	murinto murinto
Date Deposited:	26 Aug 2025 06:14
Last Modified:	26 Aug 2025 06:14
URI:	http://eprints.uad.ac.id/id/eprint/86519

Actions (login required)

View Item