Scholarlink

<p>Texture classification is a crucial field of investigation in pattern recognition due to its wide-ranging applicability in real-world tasks. In this study, we introduce a novel integration of siamese neural networks and dissimilarity for enhanced texture identification. Unlike traditional dissimilarity approaches that rely on static distance functions (e.g., Euclidean, Manhattan, cosine similarity), our method learns a task-specific dissimilarity function through joint training and triplet loss within a siamese neural network. This adaptive approach allows for a more discriminative feature space, improving robustness in problems with many overlapping classes and limited samples per class. The approach was evaluated on three texture datasets, achieving 99.6% accuracy on the Forest Species Database (FSD), 74.2% on T1K+, and 73.1% on the Describable Texture Dataset (DTD). The results are comparable to state-of-the-art methods on DTD and T1K+ and surpass the state-of-the-art on FSD.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Triplet dissimilarity: a texture classification approach using dissimilarity and siamese networks

  • 期刊论文

  • 发表日期：

    2025年6月1日

查看全文
  • Lucas O. Teixeira,
  • Diego Bertolini,
  • Luiz S. Oliveira,
  • George D. C. Cavalcanti,
  • Yandre M. G. Costa

摘要

Texture classification is a crucial field of investigation in pattern recognition due to its wide-ranging applicability in real-world tasks. In this study, we introduce a novel integration of siamese neural networks and dissimilarity for enhanced texture identification. Unlike traditional dissimilarity approaches that rely on static distance functions (e.g., Euclidean, Manhattan, cosine similarity), our method learns a task-specific dissimilarity function through joint training and triplet loss within a siamese neural network. This adaptive approach allows for a more discriminative feature space, improving robustness in problems with many overlapping classes and limited samples per class. The approach was evaluated on three texture datasets, achieving 99.6% accuracy on the Forest Species Database (FSD), 74.2% on T1K+, and 73.1% on the Describable Texture Dataset (DTD). The results are comparable to state-of-the-art methods on DTD and T1K+ and surpass the state-of-the-art on FSD.